The rest

Changelog

What shipped, in the order it shipped. Published builds only.

This tracks released builds: Windows, and macOS from 1.3.57. Work merged after a release is not listed until a build carrying it is published.

#1.3.63

16 October 2026

Review a teammate's work before it lands. A companion shared for Remote work can now work in a copy of your folder: you read what it changed under Team, Companions and apply it to your folder, or throw it away. It is on by default for a new Remote share. You can put any Remote turn back with one button, and copy its patch for a developer. Shared companions can no longer bring programs in or send them out.

Ready for a paid Pragma. Pragma sends its installation's random ID with update checks, so that when Pragma goes on sale, a copy installed before then keeps its updates free. Nothing changes for you today: Pragma still downloads and updates for free, and it always opens, licence or not. When it is sold it is a one-time purchase, $99 at the beta price and $199 after, with no subscription.

This release carries everything in 1.3.63-beta.1. Available for Windows, Apple Silicon and Intel Mac.

#1.3.63-beta.1

Pragma Beta, 16 October 2026

Review a teammate's work before it lands. Under Members and invites, a companion shared for Remote work now has "Review before it lands in my folder", on by default for a new share (older shares keep working in the folder until you choose). The teammate's companion then works in a copy of your folder, and nothing reaches your real folder until you read what it changed under Team, Companions and press "Apply to my folder", or throw it away. Applying holds back files that start code by themselves and programs, and never overwrites a file you changed in the meantime. It also closes the gap for Codex companions: they can only write inside the copy. A folder over 500 MB cannot be copied, and its turns are refused while review is on.

Put a Remote turn back. When a teammate works in a folder through one of your companions, Pragma now keeps a copy of the folder from before each turn. Under Team, Companions, "What Remote work changed" you see what was added, changed and removed, can read the patch, copy it for a developer to apply and push with their own credentials, or put the folder back with one button. A file you changed since is left alone and listed. It works for every engine, Codex included. It cannot restore links, files over 5 MB, or a git commit the companion made.

Shared companions cannot bring programs in or send them out. Programs and scripts a double click starts (and programs under another name) are refused on upload, from the outbox, and by the team service. A Remote turn is refused writes to git hooks, CI, editor tasks, .env, .npmrc, Makefile, Dockerfile and node_modules, and git is held to the commands that work on the files in front of it. After each turn you get a note of what changed and what to look at first.

Pragma in Slack, getting ready. Each person in a Slack workspace will link their own Pragma with a one-time code, so what they write is answered on their own computer and nowhere else. The Slack card says Coming soon until it is switched on; there is nothing to do yet.

This beta carries everything in 1.3.61-beta.13 and 1.3.62. The team service it needs is already live. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.62

16 October 2026

The biggest update since 1.3.60: everything Pragma Beta testers have been using, now for everyone. The Pragma Beta entries below list each change in detail.

Team. A new Agents and Team switch sits at the top of the sidebar. Team is chat in channels and direct messages, a task board with a timeline, files, and invites by email, link or code. Pip, a small assistant, keeps the board for you: ask it to make, change, give out, comment on and flag tasks, and to start channels. You can connect your own ChatGPT dot to a team as its assistant, and share a companion so teammates can ask it from a channel, or work in one folder of your computer while you are away. Roles go beyond owner, admin and member: Developer / AI Engineer, Quality Assurance and Client, and Pip, the dot and shared companions act only as far as the person asking is allowed.

Companions that work as one. A companion can run on as many models as it needs, each for the part of the job it is best at. Every companion starts with a position, such as Chief of Staff, Film Maker, Writer or Developer, with 114 more skills for them. There is one Chief of Staff for your whole account, pinned at the top, and the sidebar sorts companions by category.

Work that is not lost. If Pragma closes, updates or crashes while a companion is working, it picks the work up again when you reopen it. You can @mention another companion in a chat, and an unread count shows on every companion, room, channel and direct message, and on the taskbar or dock icon.

Simpler to use. One Show advanced options switch hides keys, folders and commands until you want them. Files a companion makes open in a panel inside Pragma. A chime sounds when an agent finishes or needs you while you are somewhere else. Companions you share, and the places they work, are easier to see.

Slack. Connect your own Slack app under Connected apps (Pragma Tools) so a companion can read, search and post in your workspace when you ask. Messaging a companion from Slack is coming soon.

Smaller things. Organize companion chats into sidebar groups, plainer messages when an engine is not signed in, faster Team, cheaper messages for some engines, fewer and better-chosen models in the picker, and more ready-made gear.

Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.13

Pragma Beta, 16 October 2026

More roles on a team. Besides owner, admin and member there are now Developer / AI Engineer, Quality Assurance and Client. Pick one when you invite someone, or change it later under Members and invites (the owner does that). A card on that page lists what each role can do. Pip, your team's dot and the team's companions act for whoever asks them, so they can do only what that person's role allows. A client can ask Pip and the dot for tasks and for a new channel, but never to delete anything, and cannot put the team's companions to work. None of them can delete a channel, remove a person or change a role.

What a channel is for. A channel can have a short description, shown under its name. Whoever can edit the channel can write it (the person who made it, an admin or a developer), and so can Pip, the dot or a companion when one of them asks.

Pip and your dot can start channels. Ask either one to make a channel, say what it is for, and put the team's companions in it. They can also put a companion into a channel you can edit.

Companions can work the board. A teammate's companion that answers you can now comment on a task, move it, mark it done and start a channel, as you. Asked to "finish T-0003", it can say what it did on the task and close it, even when it was shared with Folder only access: this changes the team's board, not its owner's computer. It cannot delete anything.

A count on every conversation. Every channel and direct message shows how many messages are waiting, not only the ones that name you. A number in red names you; a grey one does not.

A count on the app's icon. The taskbar button on Windows and the dock icon on a Mac show the number of unread messages across your companions, rooms and teams, and it clears as you read them.

Pictures. Pip's picture is now a headshot of Pip. You can choose a picture for your team's dot under "Your dot". Your own picture is one picture: changing it in Team changes the one beside your name in the sidebar, and the other way round, and Settings, Profile now has a button to change it.

Slack tools are called Pragma Tools. The Slack app you create under Connected apps is now named Pragma Tools, so it can never be mistaken for Pragma in Slack. Its setup screen says it is Slack as a tool for your companions and cannot be chatted with.

This beta carries everything in 1.3.61-beta.12. The team service it needs is already live. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.12

Pragma Beta, 15 October 2026

Slack, connected directly. Open Connected apps and add "Slack (your own app)". Pragma opens Slack with a ready-made app: pick your workspace, install it, and paste the Bot token (and, if you want search, the User token) back into Pragma. Because the app lives in your own workspace, Slack applies its normal limits, not the strict ones it puts on shared apps. Your companions can then list channels, read a channel or a thread, search, find a person, post, reply in a thread and react. The tokens stay on your computer, encrypted, and are sent only to Slack. This is new, so tell us what breaks.

Reply to a message in a team chat. Hover over a message and press Reply. The message shows what it answers, and pressing that quote takes you to the original.

Answer by name. Write "Scout, can you check this?" or "Pip, what do you think?" with no @, and the companion or Pip answers.

Change your own name. Under Members and invites you can change the name your team sees, next to your picture.

A chime when an agent finishes or needs you. A short sound when you are in another window or another chat, so a run that ends does not go unnoticed. One switch under Settings turns it off.

This beta carries everything in 1.3.61-beta.11. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.11

Pragma Beta, 15 October 2026

A companion can run on as many models as it needs. Open the models button under the typing bar, add a second model, then another, up to five besides the bot's own. For each one pick what it is for: coding, planning, reviewing, research, writing, design, debugging, testing, data, visuals, summarizing, translation, math, security, support or quick tasks, or type your own. The bot's own model leads and hands each part of the work to the model it names, in the same conversation, folder and tools; when that model finishes, the work goes back to the lead. Bots with two models work as before.

A timeline for tasks. Under Tasks, switch from Board to Timeline to see six weeks at a glance: a bar for each task from the day it was made to its due date, today marked, late tasks in red, and the tasks with no date listed below. Press a bar to open the task.

Pip answers again. Pip could not reach its model because of a mistake in how the service called it, whatever key a team used. That is fixed on the service itself.

This beta carries everything in 1.3.61-beta.10. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.10

Pragma Beta, 15 October 2026

Pictures. People, companions and the team's dot now show a picture instead of a letter. Your companions' pictures go up by themselves, yours comes from the one you chose for Pragma or one you pick under Members and invites, and Pip's is bigger and easier to see.

An Agents channel, and who is in each channel. Every team has an Agents channel with every shared companion in it, and #general has them too; neither can be changed or deleted. Under each channel's name you now see everyone in it as a row of pictures, and "Add or remove" lets whoever made the channel, or an owner or admin, choose the people and the companions. A companion only answers a mention in a channel it is in. A public channel is open to the whole team; make it private to choose who is in it. New channels can start private.

Companions answer as themselves. When a companion answers, the message shows the companion's name and picture with a small "Kyle's companion" tag, not Kyle's name.

Mentions work from the keyboard. Type @nu and press Enter or Tab to take the highlighted name; the arrow keys move between them. Only companions that are in the channel are offered.

Pictures you send are small. A picture in the chat is a thumbnail that opens full size when you press it.

Tasks. Attach files of any kind to a task, add people or companions to work on it with whoever it is assigned to, and see them on the card. A companion you add is asked in the Agents channel, and its answer comes back as a comment on the task. The buttons at the top of the board and the task's own close button are clickable again on Windows.

Roles. Members and invites now says what an owner, an admin and a member can each do.

Plainer problems. A companion whose engine is not signed in says so in plain words, with what to do about it. When Pip cannot reach its model it says why (the key, the model, the free limit or a slow answer) instead of "try again".

The dot. The team's dot is asked to look up its waiting request itself when ChatGPT's notification carries no details, you see that it is working while you wait, and a question it has not answered in five minutes is answered by Pip.

Remote work for shared companions. Share a companion with Remote work to let named people work in one folder on your computer while you are away. File reads and edits in that folder and the programs you tick are approved ahead of time; anything else Pragma is asked about is refused. Files a teammate sends land in a team-inbox folder, and files the companion saves in team-outbox come back with its answer. Pragma offers to keep the computer awake. A new Companions page lists what the team has shared. On a Codex companion the folder is held by Codex's own sandbox, which keeps writes inside it but does not stop reading other files, and the panel says so.

This beta carries everything in 1.3.61-beta.9. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.9

Pragma Beta, 14 October 2026

Send files in Team. Attach a file with the paperclip, drop it on the message box, or paste it. Pictures show right in the conversation and everything else is a card you click to save. A file can be up to 10 MB, and a team has 1 GB in all. Only people who can read the conversation can open its files, and deleting a message or a team deletes its files too.

Pip and your dot are pinned. At the top of the sidebar, above Today, Pip is always there, and so is the dot you connected. Open Pip and, if your team has no Groq key yet, a card there lets an owner or admin paste one. Open your dot and it walks you through connecting it.

Faster Team. Sending a message, opening a conversation and checking for new messages now make far fewer round trips to the service, which was the reason sending felt slow. How fast your dot answers still depends on ChatGPT, which can take a minute or two.

This beta carries everything in 1.3.61-beta.8. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.8

Pragma Beta, 14 October 2026

Connect your own ChatGPT dot to a team. If your ChatGPT plan has dots, an owner or admin can add Pragma to ChatGPT as a plugin (Team, Members and invites, Your dot shows the address and the steps), sign in to Pragma Team, pick the team, and give the dot a name. Once the dot passes a test you send from Pragma, it becomes that team's assistant under that name: people mention it with @ or write to it in their own conversation with it, and it can answer, and make, change and comment on tasks for whoever asked. It cannot delete tasks or change who is in the team. Pip stays as the fallback, and you can still ask Pip with @pip. You can pause or disconnect the dot at any time. What people ask the dot goes to your ChatGPT.

This is new and has been tested against our own copy of how ChatGPT connects, not yet against ChatGPT itself. Whether a dot can install a plugin that is not yet in the directory depends on your plan, so expect rough edges and tell us where it stops.

Pip thinks with your team's own free Groq key. Pip is a small assistant that runs in Groq's cloud, not on your computer, so Groq's free plan is plenty. An owner or admin gets a key at console.groq.com/keys and pastes it in Members and invites, under Pip. Until then Pip tells people how, with a button that opens the sign-up page. The key is checked with Groq, kept sealed, and only its last four characters are ever shown.

This beta carries everything in 1.3.61-beta.7. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.7

Pragma Beta, 14 October 2026

Team is complete. There is a task board with five columns: drag a card to move it, open one to see its brief and a ready-made prompt for an agent, add comments, flag it, or send it to one of your own companions. Search finds tasks and messages, and any message can become a task.

Pip is your team's assistant. Pip wears the mascot and keeps the board: ask it in its own chat or with @pip to make, change, assign, comment on, flag or delete tasks. It writes the brief and the prompt for you, and it always asks before deleting. Pip also acts like a teammate. It answers when you reply to it, joins a conversation now and then, starts one when the channel has gone quiet, and tells you when a task of yours is due or overdue. It reads public channels only. An owner or admin chooses Chatty, Quiet or Only when asked under Members and invites.

Invite people by email, a link or a code. An emailed invite works only for that address. A link opens a small page that opens Pragma with the code filled in, and you can paste a whole link into the join box.

Companions in the team. Share one of your companions and teammates can @mention it in a channel. It answers on your computer with its own model, so your plan pays for the tokens, and what was asked also shows in its own chat. An owner or admin can add a companion for the whole team. Folder only is an instruction to the companion, not a lock.

Know what needs you. The count on the Team button covers mentions, direct messages and tasks given to you. A desktop alert appears when it grows while Pragma is behind another window, and an optional email, never containing message text, reaches you if you have been away. One click stops it.

Pragma finds what it needs on its own. Settings, Engines, On this computer lists the programs (Git, Node, Python, FFmpeg, Blender and others) and the AI models already downloaded (Ollama, LM Studio, Pragma's own). Companions are told, so they stop asking you to install something that is already there. Blender installed in an unusual place, such as another drive, is now found.

Also fixed: on Windows the Team screens no longer sit under the minimize and close buttons, and Channels and Direct messages are separate sections.

Pip uses Cloudflare's own model for now. Team has no web version yet. This beta carries everything in 1.3.61-beta.6. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.6

Pragma Beta, 13 October 2026

Team chat is live. The Team side of the new Agents and Team switch now works for real. Sign in with your email (a 6-digit code, no password), make a team or join one with an invite code, and talk in channels and direct messages. Type @ to mention someone. A count on the Team button shows what needs you, such as a mention or a direct message, even while you are on the Agents side.

Invite codes work once, last a week, and are shown only when you make them. The owner and admins can invite, change roles and remove people. Everyone signs in with their own account: nothing is shared.

Tasks and Pip are not live yet. The Tasks entry still shows the sample board, and Look around first shows a sample team.

This beta carries everything in 1.3.61-beta.5. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.5

Pragma Beta, 13 October 2026

Files open inside Pragma. Clicking a file a companion made, in its reply or on its file card, now opens it in a panel beside the chat instead of another program. It shows text, code and documents, sheets as a table, web pages, pictures, PDFs, sound and video. A kind it cannot show, such as a Word file or a deck, says so and offers to open it in the program that can.

A first look at Team. A new Agents and Team switch sits at the top of the sidebar. Team is where people will chat in channels, share a task board, flag things and ask Pip to create, assign and update tasks. This build has a clickable preview with sample people and tasks so the layout can be judged. Nothing in it is saved or sent yet, and it is not connected to anyone.

This beta carries everything in 1.3.61-beta.4. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.4

Pragma Beta, 11 October 2026

Every companion has a position. A new companion starts by asking what it is for: Chief of Staff, 3D Blender Film Maker, Video Producer, Writer, Designer, Game Maker, Web Builder, Developer, Inbox Manager, Executive Assistant, Project Manager, Customer Support, Documents Specialist, Bookkeeper, Researcher, Marketing Lead, Sales Researcher, Data Analyst, Watcher, or your own. The position gives it starting instructions, three first things to try, and the skills it reaches for first. Companions you already have keep working as they are, with a small offer to pick a position.

One Chief of Staff, pinned at the top. The Chief of Staff is now one companion for your whole account. It sits in its own section at the top of the sidebar and can brief, schedule and coordinate every other companion. Where you had one in more than one section, the first keeps the role and the others carry on as ordinary companions.

The sidebar sorts by category. Companions with a position are grouped under Create, Build, Run the day and Grow. A team or a group you made yourself still wins.

3D Blender Film Maker. It shows whether Blender is on this computer and offers the download. Sound effects come from ElevenLabs when you have saved a key (Agent profile, Voice), and are made in code when you have not. If Codex and Claude are both signed in it offers to use Codex for the look and Claude for the build.

Companions schedule their own work. Tell a companion you want a briefing every weekday at 7pm and it asks what it needs, then sets the routine up itself instead of sending you to the Tasks page.

114 more skills for the positions, from film and animation to marketing, sales, design, websites and games.

This beta carries everything in 1.3.61-beta.3. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.3

Pragma Beta, 11 October 2026

A new companion greets you by the name you gave it. Naming a companion on its setup card used to leave its first line saying the name it was made with. The line now carries the name you chose.

It speaks first after setup. When you finish the setup card and the companion's engine is signed in, the companion follows up with the one thing it needs to know to do the job, for example which days a daily brief means. If the engine is not signed in or not installed, you only get the corrected greeting.

This beta carries everything in 1.3.61-beta.2. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.2

Pragma Beta, 6 October 2026

Work is not lost when Pragma closes. If Pragma is closed, updated or crashes while a companion is working, it picks the work up again when you reopen it, in a chat or in a room, and says so in the chat. A companion whose engine stops partway is also picked up again, up to twice.

@mention a companion. Type @ in a chat to pick another companion. It answers in that chat under its own name, with the last few messages along, and several can answer at once. Not in shared chats yet.

Simpler by default. One Show advanced options switch in Settings hides keys, folders, commands, addresses, webhooks and program paths until you want them. The computer panel no longer shows the Box key or Runs on; they sit under Advanced.

Fewer models. The model picker opens on the best, the everyday and the fast model. GPT-5.5 is no longer offered, Pragma's own local models are listed once, and the list follows an engine update when you come back to Pragma.

Add gear by talking. Ask a companion to add a tool and it finds it, registers it and runs it. Every ready-made gear shows an example of its result before the first run, with a Try once button, and Gear has an icon of its own.

Better website starting points. References replace Templates: seven finished live pages, one for each style, with real 3D, and twelve parts to borrow. The check that catches a page scrolling sideways on a phone is stricter, and a gallery that did not show most of its cards is fixed.

Cheaper to talk to. Antigravity and pi get Pragma's setup once per conversation instead of on every message, and a plain hello carries about a third less setup. Skills load one at a time, and the Skills page tells you when one did not load.

Smaller fixes. Delete confirmations are Pragma's own dialog everywhere. The companion menu's Delete is no longer cut off in a short window. The setup card of a new companion takes typing. Deleting a website no longer ends in a raw error, and Clear on the working folder works. Pragma's own rules now reach every companion.

Slack, not switched on yet. The Slack card is in this build, but the Slack service is not live, so Add to Slack will say it is unavailable for now. Listening is coming.

This beta carries everything in 1.3.61-beta.1. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61-beta.1

Pragma Beta, 2 October 2026

Organize companion chats. Create sidebar groups from Organize below search, choose the conversations to include, and rename or remove groups without changing their teams or losing their history. Existing channels can be grouped alongside companion chats.

Collapse the clutter. Each heading has an expand/collapse control, and Collapse all and Expand all remember your preference. Search temporarily shows matching chats. Folder markers and Edit buttons distinguish your chat groups from functional teams.

This beta carries the same features as stable 1.3.61. Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.61

Stable, 2 October 2026

Organize your chats. Use Organize below search to create sidebar groups such as Work, Personal or Clients. Choose the companion chats and existing channels to keep together, rename a group, or remove its grouping while keeping every conversation. Your companions' teams and permissions stay the same.

Keep the sidebar tidy. Collapse any heading, or use Collapse all and Expand all. Pragma remembers what you collapsed, and search still reveals matching chats. Chat groups have their own folder marker and Edit button.

Local models and connections. This stable update also includes the fixes from beta.17: context sizing for expert local models that use system memory, retained tools and task instructions during long-chat recovery, simpler Connect actions, current Kimi credential detection, conversation filters and clearer phone access status.

Payments remain disabled. Available for Windows, Apple Silicon and Intel Mac.

#1.3.60-beta.17

Pragma Beta, 2 October 2026

Local models keep their tools. Expert models that use system memory now get a context window sized for the memory available on your computer. When a long conversation needs trimming, companions keep their file tools, skills and task instructions.

Connect from one place. Engine rows show whether they are ready and offer Connect. Technical settings stay under Advanced. Kimi sign-in recognises the current credential files, and Windows finds its usual installation folder. Clearing a custom CLI path restores automatic detection.

Find the right conversation. Filter the sidebar by Companions, Rooms or Team chats, with a separate section for internal team conversations.

Phone access has a clear status. The switch shows when access is turning on or off and keeps the chosen state when an older status check arrives.

Payments remain disabled. This update is available on the beta channel for Windows and macOS.

#1.3.60

Stable 1.3.60 and Pragma Beta 1.3.60-beta.16, 2 October 2026. Earlier beta builds ran from 28 September to 1 October.

Gear is easier to set up. Ready-made gear uses simpler cards and practical setup fields. JEV keys and technical scoring choices no longer appear in customer setup. A connected companion can help with the work; managed scoring is still being prepared.

Websites have previews. Website cards show a preview, their current status and clear actions to open, share or publish. Templates give you a starting point, and editing keeps the website's working folder attached.

Voice and setup are clearer. Choose your microphone, keep an unsent transcription and pause for up to ten seconds while speaking. Engine setup offers guided installation and sign-in, with technical choices under Advanced. Work shows elapsed time and clearer progress, and stopping a Claude task now reports it as interrupted.

Community videos recover better. The player loads protected media in chunks, supports range requests and gives a clearer retry path when a video cannot load.

Payments remain off. This release excludes payment checkout and subscription management. Managed scoring, always-on cloud execution and custom-domain provisioning still require service activation and are not enabled by this update. Existing local routines continue to run while Pragma is open.

Pragma Beta. New features now go to Pragma Beta before everyone else. It is a separate app that installs beside Pragma, with its own bots, chats and settings, so testing it never touches your real work. It has its own icon and says Beta in the corner. Pragma Beta is for alpha testers, who each get their own download link by email. Pragma itself now gets fewer, tested updates.

Meet Pip. A new Pragma starts with Pip, the Pragma mascot: a soft charcoal egg with an orange band round his middle, animated in 3D. Pip asks a few questions (your name, whether Pragma is for you or a business, what you do, which jobs to hand off, where your work lives) and then builds the team your answers describe: bots of your own, or a business team with its own channel and Pragma Setup to connect your apps and schedule the work. You can meet Pip again from Settings. (Pip took over from Grid in beta.4.)

Companions. Every bot is now one of sixteen companions, plush felt characters in a round profile picture: Mallow the bunny, Fenn the fox, Nori the snail, Mote the butterfly, Luma the star, Nimbus the cloud, Tuck, Riff, Pebble the penguin, Puck the frog, Kiko the axolotl, Otis the otter, Bumi the mushroom, Juno the jellyfish, Bao the dumpling and Oona the moon. Each sits on a circle in its bot's color. They wave hello when they arrive, bob along while they work, and breathe when you point at them; everywhere else they hold still, so a long sidebar costs nothing. Pick a bot's companion and color in its profile. Chats and the sidebar show a headshot; the profile shows the whole companion. Bots you already have get the companion nearest their old shape, and new bots are spread across the cast. (beta.4)

Bots keep going when a chat gets long. A conversation that grew past what the bot's model can hold used to stop with an error. Now the bot trims what it sends and tries again, and if the turn still does not fit, hands it to a model that can hold it, so the work carries on. (beta.4)

Opens in a second. On some computers Pragma showed a blank window for up to 20 seconds while it started. The window now appears straight away and Pragma finishes starting behind it: usable in about 2 seconds in testing. (beta.4)

See the cost while it works. While a companion works, a quiet line under it counts the tokens used so far and, where your plan reports it, how much of your limit that is, so a long job no longer looks silent. Pictures and files it makes along the way show their cost too: what the turn has used so far, then the final figures when it finishes, and for a picture, what making the picture itself took. (beta.14)

Websites with 3D and real motion, by default. Ask for a website and your companion first asks a few questions in one card: what the business is (with real details to use word for word), websites you like, the visual style (eight looks to pick by eye), how much animation, what visitors should do, what you already have, and where to save it. Every question already has Pragma's choice selected, so you can press Start building straight away, and saying "no questions" still skips it. It then builds from a kit of tested pieces: glossy 3D objects, glass and plates, photos with depth, embers and steam, product spins, smooth scrolling, pinned galleries and masked reveals. It checks the result in a real browser, on a phone too, before it shows you, and a second model designs while the first builds. (beta.15)

Optional Blender for richer 3D. On a computer that can run it, Pragma can offer Blender so your companion can model and light a hero object (a wine glass, a plate, a bottle, a logo) for your website. It shows a card with the download size and where it goes, and nothing downloads until you press the button. Without Blender, the built-in 3D is used and the site is just as complete. Settings, Engines, 3D tools. (beta.15)

Voice on Windows, and voice you can see. Dictation and calls now work on Windows, with speech turned into text on your computer: the first time, Pragma offers a one-time download of 190 MB (60 MB on slower computers). A bar above the message box now shows when the microphone is on, with a live waveform, a timer, and buttons to keep or throw away what you said, and recording stops by itself after a pause. On the Mac it no longer keeps listening after you send, switch chats or quit. Voice mode shows Listening, Thinking or Speaking, and companions can speak with your computer's own voices, with no key needed. (beta.15)

Deleting a website deletes its folder. After a prompt that names the folder and says it cannot be undone. Pragma refuses for a folder that is a whole drive, your home or main folders, Pragma's own data, a code project, or shared with another website, and offers to remove it from the page instead. (beta.15)

Companions keep to the job. A companion now always gets the instructions for the current task, even deep in a long conversation, and no longer picks up your own Claude Code or Codex setup (your CLAUDE.md, skills, plugins and memories) unless you turn on My Claude Code tools or My Codex tools. (beta.15)

Websites that look designed. Ask for a website, even in a few words or with "no questions asked", and your companion now starts from a real reference design, makes real photos, fills in believable content (and lists what it made up for you to correct), then looks at the site on a computer and a phone and fixes what it sees before showing you. Two companions together now do better than one: the designer sets the direction and hands it over as a file, the builder builds, and the designer checks it. Every site gets favicons in every size, a sharing image, the tags search engines read, a sitemap, a 404 page and the basics AI search engines use to find and quote a business. Improving a site now changes it in place and keeps its last 10 versions, which you can compare and restore, and publishing sends only what changed. (beta.13)

Know what it cost. Under every companion reply, beside the model's name, you now see the tokens that reply used and, where your plan reports it, how much of your limit that was. (beta.13)

Companions know what you have. Each companion now knows which engines you are signed in to, which can make pictures, which other companions it can hand work to, and which apps and gear you have, so it uses them instead of asking. Pictures are made with what you already have, such as Codex, without hunting for another key. The list of models now comes from each engine itself, so new models appear as soon as your engine offers them; if an engine is out of date, the model picker says so. (beta.13)

Thirteen more gear, twenty-seven in all. Content Repurposer, Proposal Builder, Ad Creative Tester and Onboarding Pack; Expense Report, Contract Checker, Bank Statement Matcher and Supplier Price Tracker; Daily Site Report, Form-to-Sheet Checker, File Tidy, Reading List Digest and Calendar Guard. The Gear page groups them by what they help with, and any report a gear writes opens from its card. (beta.13)

Nineteen more skills. Companions now make proper Word documents, spreadsheets with working formulas, PDFs, slides and invoices, keep simple books, analyse data, and write better emails, posts, support replies, meeting notes, plans and research, with the sums done by code rather than guessed. (beta.13)

Storage. Settings, Storage shows what Pragma keeps on this computer and offers safe clean-ups, each listing exactly what goes before you confirm. Logs and caches stay trimmed on their own, and a notice appears if Pragma grows past 5 GB. (beta.13)

Smaller things. Your own messages no longer glow. Pragma uses its own icons everywhere. The switch on a website card now clearly means this computer only; your online links keep working. (beta.13)

Websites. When a companion builds you a website, it now lands on a new Websites page. Switch it on to look at it on your own computer, share a preview link that works for three days, or put it online at a name you choose, like yourname.trypragma.site. There is no hosting to set up and nothing technical to learn: pick a name and press Publish. Publish again to update it, or take it offline whenever you like. Free accounts get three live sites of up to 50 MB each, and sites carry a small Made with Pragma badge. Putting a site online needs you to be signed in to Pragma with your email. Just ask a companion to make you a website. (beta.12)

Direction first. Ask for something big or creative, like a video, a design, a website or a long document, and your companion now shows you two or three directions before building anything: what each one is, which it would pick, and roughly what it costs. Pick one, mix them, or say something else in your own words. Set Plan first in a companion's profile to For big jobs, Always or Never. Quick requests skip it. (beta.11)

Point at your attachments. Every picture or file you attach gets a label, like image1 or file2. Type @ to pick one, and say "look at @image2 and make the button bigger". The numbers carry on through the whole chat, so you can point back at an earlier one. (beta.11)

Nine more pieces of gear. Follow-up Finder finds quotes and conversations that went quiet, ranks which are worth chasing and drafts the nudge. Invoice Chaser finds overdue invoices, adds up what is outstanding and drafts reminders that get firmer each time. Review Reply sorts reviews into happy, mixed, angry and fake-looking, flags the ones that need care, and drafts replies. Document Intake reads what you drop in a folder, works out what each file is and proposes a name and a place for it; nothing moves until you say so. Resume Screener ranks CVs against the must-haves you wrote, with names and personal details left out of the scoring; it sorts, you decide. Meeting Notes to Tasks turns a transcript into decisions and actions, and suggests each companion's actions on its to-do list. Morning Briefing puts your calendar, the emails that matter, your companions' to-do lists and yesterday's results into one short message. Timesheet Checker flags missing days, double entries and unusual overtime, and drafts a payroll register for you to approve. Website Health Check walks your own site for broken links, missing titles and descriptions and other problems, in order of what to fix first. The Gear page now has a search box and shows everything in one place. (beta.11)

Construction drawings. Share a set of construction drawing PDFs and your companion indexes it once: every sheet, its number, title, scale and type, and which sheets refer to which. After that it answers questions about the set quickly and cheaply, measures areas, lengths and counts from the drawing itself with a confidence next to every number, and drafts an estimate from your own rate sheet. Scanned sheets are marked to check by hand, no rate is ever made up, and the estimate is a draft for a person to review. (beta.11)

Companions talk like a friend. Every companion now has a way of talking: Friendly, the default, uses plain words, leads with the result and keeps it short, with no jargon; Developer is technical and exact. Switch it by right-clicking a companion or under How it talks in its profile, and choose the default for new companions in Settings. (beta.11)

You can see what things cost, and where they went. Every file, image or video a companion makes now shows how many tokens it took and, where your plan reports it, how much of your limit that was. It also shows exactly where the file was saved: click to open it, or show it in its folder, and any file path in a reply is clickable too. If a companion is about to make files and is not connected to a working folder, it asks first: connect a folder, make a new one, or use its private folder for now. (beta.11)

Unread shows how many. The sidebar shows the number of unread messages for each companion and room, not just a dot. (beta.11)

Four more pieces of gear. Inbox Triage sorts your recent email from the Gmail or Outlook you connected into hot leads, questions, bills, can wait and junk, and your companion drafts replies to the ones that matter; nothing is sent. Receipt Sorter reads a folder of receipt and invoice PDFs and photos, pulls out the vendor, date, amount and tax, sorts each into your expense categories and writes a spreadsheet, without moving or deleting your files. Competitor Watch checks your competitors' pages, tells you only about changes that matter, like a new price or offer, and your companion writes what you might do about it. Contact Finder finds the public email, phone and social links of businesses from their own websites, ranks how well each fits what you sell, and hands the best to your companion for a first message; it never guesses an email and never sends anything. Every result exports as CSV. (beta.11)

Shared companions have their own place. Conversations with people you shared a companion with now sit under Shared companions in the sidebar, not under Channels. (beta.11)

Companions keep working. Each companion now has a to-do list: work you or another companion gave it, duties it repeats, and ideas it proposes. Turn on Keep working through the list and, when it is free, it takes the next item, one at a time and within a daily limit you set. Nothing it proposes runs until you approve it, and its ideas do not ping you one by one: your Chief of Staff posts one digest a day with everything proposed and everything finished, to approve or dismiss. (beta.10)

Routines that do not get stuck. When you set up a routine, you can say what it may do on its own, like running a report command or posting in one channel. Anything else stops with a clear card, "it wanted to do this: allow for this routine, only this time, or deny", instead of waiting silently, and anything destructive still asks. Each routine shows when it runs next. Routines only run while Pragma is open and the computer is awake, so Pragma now says so, can keep the computer awake while routines are scheduled, and shows a run it missed while closed, with Run now. (beta.10)

Setting up is easier. Ask a companion to install an engine or gear Pragma already supports, even by pasting its GitHub link, and it shows an Install in Pragma card that opens the one-click setup. A companion needs only its instructions and its memory, not soul.md or other setup files, and the profile now says so, with Help me write instructions to draft them from a few questions. (beta.10)

Bring your companions into Pragma Beta. The first time you open Pragma Beta, it offers to bring your companions over from Pragma: their instructions, models, memory, rooms and settings, and their chats if you like. Pragma itself is not changed, and you can do it later from Settings. Keys and sign-ins stay in Pragma, so add them again in Beta if a companion needs one. (beta.9)

Gear: ready-made tools your companions use. The Apps page is now Gear. Besides what your bots built, it has ready-made gear you install in a click: paste any keys it asks for, run it when you like or every day, and your companions can run it for you from a chat ("use Lead Scan for plumbers in Austin"). The first piece is Lead Scan, which finds people on Reddit asking for what you sell. JEV, a very fast and cheap decision model, ranks every post in seconds for a fraction of a cent, and your companion drafts replies for only the best few; nothing is ever posted for you. JEV is invite-only, so without a JEV key Lead Scan scores the posts with your companion's cheapest model instead, and each run says what it cost and what JEV would have cost. Gear sets up the Python it needs by itself the first time. (beta.8)

Mini Companion, part two. The person you share with now sees the companion itself at the top of their chat, not a letter, and it moves while it works. If you share more than one companion with them, they tap + and paste the next link, then switch between them from the row at the top. Each stays its own private chat, and removing one leaves the others working. (beta.7)

Mini Companion: share a companion with one person. Open a bot's profile and use Mini Companion to let one person chat with that companion from their phone or computer browser, with nothing to install. Name them, choose how much the companion can do for them, pick Chat and Voice, and send the one-time link or QR code (it works once, within 24 hours). They get their own private conversation with that companion and nothing else: not your other chats, bots or settings. There are two levels. Folder only (the default) keeps the companion inside its own folder: it can read and edit files there, but anything past it, like running a command, opening a file elsewhere or going online, waits for your OK, and your connected apps, phone and the companion's memory stay out. It works on Claude and on chat-only engines; companions on other engines, or set never to ask, can only be shared with Full access for now. Full access is for people you trust: the companion can use everything it reaches for you, connected apps included. Either way, anything the companion asks approval for comes to you, however its own approval settings are set. You can read along, see each person's level under Shared with, and remove them in one tap. To change someone's level, send them a new invite. For now they need to be on your Wi-Fi or your Tailscale network. Settings, Companion is now called Full Companion: all of Pragma on your own phone. (beta.6)

Quicker to load, smoother to use. Pragma used to load every screenshot and picture in every chat before it could draw the sidebar: 7 MB on one real workspace, now under 300 KB. Pictures load as you scroll to them. Starting up no longer asks for everything twice, several checks of your engines now share one, and switching back to a chat you have already opened takes about half as long. (beta.5)

Companions look right. A moving companion no longer shows a frozen copy of itself behind it, and headshots are centred on the face without the floor the companion was standing on. (beta.5)

Hermes chats stay clean. A bot on Hermes got Pragma's setup notes (who it is, its memory, its tools) glued to every message, so the same chat opened in the Hermes app showed that long block at the top of each one. Hermes now gets them once per conversation, and again only when something in them changes, like a new note in the bot's memory. (beta.2)

Animated films and 3D characters in Blender. A new built-in skill, Blender Animation, lets a bot make short animated films in Blender, in clay, glossy CG or feature-film style, with its own music and sound. Ask for a known character from a picture and it now measures the character against your references part by part, matches each picture's camera, and sculpts heads, hands and arms as smooth shapes, so the result looks like the character instead of a stack of boxes. It works on any computer with Blender 5 installed. The Blender Asset Builder skill, for game and web 3D assets, got the same update: baked textures, painted wear and grime, rigged characters with animations, and GLB or FBX export. (beta.3)

#1.3.59

27 September 2026

Updates download in seconds. Pragma's updater read its download in a way that could slow it to a few kilobytes a second, so an update could take hours on a connection that fetches the same file in seconds anywhere else. It now downloads at your connection's full speed: the Windows update came down in under 3 seconds in testing. One catch: the update to 1.3.59 still downloads the old way, because the copy you have does the downloading. If it crawls, download Pragma from trypragma.xyz and install it over the top. Every update after this one is quick.

#1.3.58

27 September 2026

Sign in with one click. Claude and Codex now sign in from the app itself. Press Sign in to Claude (or Codex), approve in the browser that opens, and the card updates by itself when you are done. No terminal, nothing to copy or paste. If Claude's page shows a code instead, paste it into the box that appears. Engines whose sign-in asks questions in a terminal still offer one, and on a Mac that terminal now actually opens.

A set of Qwen models to choose from. Local models now come in four sizes, each marked with how it fits your computer:

• Qwen3.5 35B-A3B, the strongest and still quick. It reads only 3B of its 35B weights for each word. Needs about 25 GB: a Mac with 36 GB or more, or a PC with plenty of memory.
• Qwen3.5 27B, careful and thorough, and best on a PC with a 24 GB graphics card. It is slow on a Mac.
• Qwen3.5 9B, the one to start with on most computers.
• Qwen3.5 4B, for smaller graphics cards.

On a Mac, a model that does not fit the memory macOS gives the graphics is now marked too big instead of slow, since there is no separate memory for it to spill into.

Lighter still. Small mascots draw at 30 frames a second, and every mascot slows to 15 while Pragma is in the background.

Local models on a clean Windows PC. A PC that never installed the Visual C++ runtime can now run local models too.

#1.3.57

27 September 2026

Pragma on the Mac. Pragma now runs on Macs with Apple silicon and on Intel Macs, signed and notarized by Apple, so it opens without warnings and keeps itself up to date like the Windows app.

Local models on Apple silicon. On a Mac with an M-series chip, Pragma downloads and runs local models itself on the Mac's graphics, with nothing else to install. What fits is worked out from the memory macOS lets the graphics use.

Skill recordings transcribed on the Mac. Recording a skill on a Mac can now turn your narration into text on the Mac itself, the same way it does on Windows. The recording never leaves your computer.

Smoother with many bots. Mascots scrolled out of sight stop animating, and the ones on screen draw at most 60 frames a second and skip changes that change nothing. With a long sidebar the app does several times less work each frame, and a 120 Hz display no longer doubles it.

#1.3.56

26 September 2026

Apps. What your bots build now lives in one place: Apps, in the sidebar. An app that runs on your computer (a dashboard, a tool, a site) can be opened, started, stopped and restarted from there, whether or not a bot is working, and shows what it printed when something goes wrong. Apps you leave on come back on when Pragma starts. Bots add what they build on their own; you can add one yourself too.

See which model wrote each reply. On a bot with two models, every reply now says which model wrote it, so you can see the planning model plan and the coding model code. When the second model hands the work back, the lead model is reminded what you asked for, so it picks up where it should.

Pragma's own icons. Every icon in the app is now Pragma's own, drawn from the mascot: its speech bubble and its eyes. Settings has eye-shaped slider knobs, the phone app and a bot's computer carry its face, and the Chief of Staff wears a crown with eye jewels.

#1.3.55

26 September 2026

Auto mode stops for Windows deletes too. It always stopped for rm -rf, but PowerShell and Command Prompt spell the same thing differently, so a bot in Auto mode could delete a folder with Remove-Item -Recurse without asking. Now it asks first for Remove-Item or del with -Recurse or -Force, rd /s, del /s, formatting a drive, and shutting down or restarting the computer. Deleting one file, copying and listing still run on their own.

Memory reads as memory. When a bot saves something to its memory or reads it back, its steps now say Updated memory and Read memory instead of a bare Read and Edit. A MEMORY.md in one of your own project folders is still just a file.

A project team works in the project, everywhere. Bots made by Create project channel now also get the project folder as their own working folder, so one you message on its own works in the project too, not only in the channel.

#1.3.54

26 September 2026

Set up my business. Open Teams and press Set up my business. Pragma asks five plain questions about your work, then proposes a small team: each bot's job, the apps it uses, what it must ask you first, and when things happen on their own. Say yes and it builds the whole thing: the bots, their instructions, their connected apps and their schedules, asking you before each step. It starts from patterns for team operations in Notion and Drive, a contractor's back office, and an agency or content studio, and adapts them to what you said.

#1.3.53

26 September 2026

Alpha spots go to people using them. While you are signed in, Pragma notes once a day that it was opened: the date, nothing about what you did, kept for 120 days. The team uses it to see which alpha testers still use Pragma. After 10 days away you get one friendly reminder, and nobody loses a spot without the team deciding. The privacy policy has the details.

#1.3.52

26 September 2026

Two models per bot, each doing real work. Give a bot a second model and say what each one does, for example: this bot's own model writes the code, the second one plans. Both now work in full: they see the same conversation, work in the same folder with the same tools, and pass the work to each other when the next part is the other's job. The chat shows each handover. Before, the second model could only give a one-off answer, without seeing the conversation or your files, and it had to be Claude or Codex; now it can be any model Pragma runs, local ones included. Pick it from the pill under the typing bar, or in the agent's profile.

Your bot stays in charge. Your messages always go to the bot's own model. When the second model finishes its part, the work comes back to it by itself, so a smaller model that finishes a plan does not stop to ask whether to go on. Stop stops both, and nothing is passed on after it.

#1.3.51

26 September 2026

Obsidian. Point Pragma at your Obsidian vault in Settings, Obsidian, and every bot can search your notes, answer from them and write new ones Obsidian shows properly: links, properties, tags, and daily notes in your vault's own folder and name format. Nothing to install in Obsidian. Local models can answer from the whole vault at once.

Your notes stay yours. Bots add notes and append to them on their own, and ask before changing a note that is already there. However a change happens, Pragma keeps the previous version of any note a bot changes or deletes: Settings, Obsidian lists them, and Restore puts one back exactly as it was.

26 September 2026

Local models answer from far more than they can hold. Hand a local bot a file or a folder that is far bigger than its window, a book, a long report, a big log, a whole project, and ask about it. Pragma reads every part of it in fresh contexts on your graphics card and gives the bot one answer per question, with where in the file it came from. On War and Peace, six times what Qwen3.5 9B can hold, with five facts hidden in it that no search could find, the bot answered all five in 13 minutes; without this it found none. Text files for now: PDF and Word are next.

Local bots keep notes on long jobs. A local bot keeps the task's goal, its decisions and what it found in notes Pragma shows it on every turn, so a long job survives the model summarising its own conversation away. The notes are plain files in the bot's workspace, under tasks.

#1.3.49

25 September 2026

Local models, much faster, and they can see. On an NVIDIA graphics card Pragma now runs local models on CUDA, which on Windows keeps the whole model on the card: on an RTX 5070, Qwen3.5 9B went from 7 words a second to 93. The context window is sized to what your card has free, and Pragma says so when other apps leave too little room. Qwen3.5 9B and 4B can now look at images you send them. A model you downloaded before gets the faster runtime and its vision file in the background and uses them from its next start.

Manual, Auto or Bypass. The switch under the typing bar is now a menu with three modes. Manual asks before each step. Auto approves routine steps and stops for anything destructive or touching passwords and keys. Bypass approves everything; questions it asks you still come to you.

Your second model, one click away. A partner pill sits next to the mode: pick the partner model and split the jobs, say "this bot codes, the partner plans", in two clicks.

Send now, on every engine. A queued message has a Send now button whatever the bot runs on. Where the engine cannot take a message mid-task, Send now stops the current step and sends yours next, with everything so far still in the conversation.

New bots ask what they are for. A new bot's first chat asks for its name, purpose and goal, and it reads them at the start of every conversation.

Ready-made teams work again. The six ready-made teams on the Teams screen now come with the app, so they load without a connection.

#1.3.48

25 September 2026

No second copies of skills you already have. Bring in a skill, or a whole library, and Pragma first checks whether any of it is already built in. If some is, it says which and offers to skip those and import the rest. Two copies of one skill compete with each other and can load twice in one reply; skipping keeps the one Pragma updates with each release. You can still import yours anyway, and it now installs beside Pragma's instead of disappearing behind it.

#1.3.47

25 September 2026

Two models, one bot. Pair a bot with a partner model and tell it, in your own words, how to split the work: "Astra plans and reviews, Opus writes the code". The bot asks its partner mid-task for its share and carries on, in the same conversation, with no second bot to start or wait for. The conversation shows who was asked and how long it took. Set it up in the bot's profile, under Partner model.

A calmer conversation. While a bot works you see one line: what it is doing, in its own words, and about how long is left, counting down. The individual steps fold into "Worked through 14 steps", one click away when you want them.

Bypass permissions, where you type. A switch under the typing bar lets a bot work without asking you to approve each step. It still stops for anything destructive.

Queue with Ctrl+Enter. While a bot works, Enter adds your message to the task in progress; Ctrl+Enter (Cmd+Enter on a Mac) holds it for when the task is done.

Videos play in the chat. A video a bot makes plays right in the conversation, with seeking. Pictures show at a sensible size and open full size with a click.

Your brand kit. Settings, Brand holds your brand name, colours, fonts, notes, brand guide, logos, font files and inspiration: posters and designs you made or want to imitate. Every bot knows where it is and follows it for anything it designs for your brand, with your real logo files and fonts.

#1.3.46

25 September 2026

Remake a film you like. Give a bot a film and it rebuilds it shot for shot for your product, locked to the film's own soundtrack, with real 3D shots made in Blender and a finish at 60 frames a second and 4K. It checks its own work against the original frame by frame before handing it over.

Posters and graphics with real type. Ask for a poster, flyer, Instagram post, story, thumbnail or link preview and the bot designs it as a real layout, with real fonts in your colours, at the exact size it will run, checks it for overlaps and readable contrast, and hands you the PNG and a sharper 2x copy. Words are set as type, never painted in by an image model.

Videos of every kind. Launch and promo videos from a website or a brief, explainers, motion graphics, captions, talking-head recuts and slideshows.

Bots that show you how. Ask for a video or a graphic and the bot first explains how it will go, what it needs from you and what you can ask for, and says so plainly when a stronger model would do the job better. If something the work needs is missing from your computer, such as Blender, Node.js or FFmpeg, it tells you what it is for and where to download it, and installs it only if you say yes.

Your work stays yours. Nothing about the videos and graphics you make is sent anywhere without your yes.

#1.3.45

25 September 2026

Usage analytics. Pragma now tells us which features get used, so we can see what to fix and build next for the alpha: the app opening, signing in, the first-run steps, and actions like creating a bot or sending a message, linked to the email you signed in with. It never sends your conversations, prompts, files, bot replies or a recording of your screen. You can turn it off in Settings, General, Usage analytics, and then the app does not contact the analytics service at all. The privacy policy has the details.

#1.3.44

25 September 2026

A community for the alpha. Community, in the menu behind your name, opens the Pragma community in your browser: a forum and chat rooms where alpha testers post bugs, ideas and questions, share what they have built, and talk to each other and to us. Anybody can sign in and read; testers post.

#1.3.43

24 September 2026

Mascots with feelings you can see. Each mascot's face now has brows, cheeks and a mouth that change with its mood. Its brows jump when it is surprised, knit when it is cross and tilt up when it is sad; its cheeks go pink when it is happy, shy or excited; and it grins, gasps or wobbles when the moment calls for it. Small signs help the mood read at a glance, even in the sidebar: z's float up while it sleeps, a tear rolls down when it is sad, it sweats when it works hard or gets scared, sparkles appear when it celebrates, three dots rise while it thinks, and an anger mark pops up when it is cross.

Livelier moves. Moods arrive with a little bounce instead of simply switching, and an idle mascot blinks more often, sometimes twice in a row.

A close button you can reach. On Windows, the close button on a bot's profile and on the inspector no longer hides under the window's minimise, maximise and close buttons.

#1.3.42

24 September 2026

A first run that matches the app. Signing in, the welcome, the check of your engines and the permissions step now share one look with the rest of Pragma: one card, one kind of heading, the same fields and buttons, and a marker for which step you are on. The screen that helps you install an engine matches too.

Small fixes. A problem signing in shows in red rather than orange, and each engine on the engines step shows its version number.

#1.3.41

24 September 2026

Settings that match the rest of the app. Every section of Settings, a bot's profile, its computer panel and the inspector now share one set of controls: the same cards, buttons, text fields, switches and choice rows, under the same kind of header. Nothing moved or went away; it all works as before and simply looks like one app.

Small fixes. The skin picker shows all five skins in one row, and an engine without its own logo no longer pushes its name across the Engines list.

#1.3.40

24 September 2026

A calmer window. The top of a conversation now shows who you are talking to, the task you are in, the model and two icons. What the task has used, its working folder, the bot's computer, the execution timeline and the inspector are all in the More menu (the three dots). While a bot works, its typing bubble says what it is doing. Channels work the same way: who answers a plain message is a dropdown under the channel's name.

A tidier sidebar. Settings, updates, a chat with us and how the sidebar looks are behind your name at the bottom, next to icons for Tasks & routines, Teach a skill and Connected apps.

The mascots come alive. Their eyes follow your pointer anywhere in the window, they say hello when you point at them, they watch you while you type, and their bounces and breathing are big enough to see in the sidebar.

The newest Claude models. Claude Opus 5.5 and Claude Fable 5.1 are in the model picker.

Dark skins fit the window. On Windows, the minimise, maximise and close buttons now match a dark skin instead of sitting on a white block.

#1.3.39

24 September 2026

A new look. Pragma is redesigned from the ground up. Each conversation sits on its own clean sheet beside the sidebar, with a clearer header, new message bubbles and a new message box. Small animations show where you are: the highlight glides to the bot you pick, panels slide open, and the send button answers your click. All five skins carry the new design, and the Inter typeface now ships with the app, so text looks the same on every PC instead of falling back to a system font.

Deleting asks first. Deleting a bot, a channel or a task now asks before anything goes, because its conversation goes with it and cannot be brought back. Archive still takes one click and can be undone.

Nothing you write gets lost. A message that fails to send goes back into the box instead of disappearing, and something that fails to delete comes back into the list.

Approvals answer once. A double click on Allow no longer sends two answers, and a request that ended because you cancelled the turn now says so instead of showing Denied.

Fewer surprises. Changing a setting that restarts every engine warns you first while bots are working, since it stops what they are doing. Check for updates no longer says "up to date" before it has checked. The bot and channel menus open from the keyboard with Shift+F10 or the menu key, and the arrow keys move through them.

Fixes throughout. Codex no longer leaves a turn hanging when it fails to start, Pi says why it stopped, and a long command waiting for your approval is shown in full. Routines, webhooks, plugins, the team and skill libraries, GoHighLevel setup, voice and local models all got fixes too.

#1.3.38

18 September 2026

An image a bot makes shows up in the chat, wherever it saved it. Until now only pictures written inside the bot's own folder appeared. A bot that renders with its own script or pipeline, and then lists the files it wrote, now has those pictures shown in the conversation as well. Up to four per reply, and bots are told to include the path so this happens on its own.

#1.3.37

18 September 2026

Cancel a message you sent while a bot was working. Each waiting message now has its own row above the box, with an X to drop it. A cancelled message is never sent and never reaches the bot.

Send a waiting message into the turn that is already running. Where the engine supports it, the row has Send now: the bot picks the message up before its next step, inside the turn it is already on, instead of after it finishes. Engines that cannot do this say so and leave the message queued.

#1.3.36

17 September 2026

Images a bot makes show up in the chat. A bot on Antigravity that generated an image used to save it in Antigravity's own folder and reply with a link that would not open. The picture now appears in the conversation, and it is saved in that bot's workspace under images, so you can open it or use it again.

Links to files open. When a bot links to a file on your computer, such as a document it wrote or an image it saved, clicking the link opens it. Programs and scripts are shown in their folder instead of being run.

Every generated image is kept as a file. Images from Pragma's own image generator are saved in the bot's workspace too, not only shown in the chat, and the bot is told where.

Sign-in codes arrive reliably. Sign-in and licence emails now come from hello@trypragma.xyz. Replying still reaches us.

#1.3.35

17 September 2026

Sign in with your email. Pragma now asks for your email address and sends you a 6-digit code. There is no separate sign-up: the first time you sign in creates your account. Your connected apps belong to that account, so they follow you through updates, a reinstall or a new computer. If you were signed in before, sign in with your email once and the apps you had connected come with you.

Chat with us. The button at the bottom of the sidebar now opens a WhatsApp chat with Systo, for questions and help.

Pragma is sold through Paddle. New purchases are processed by Paddle.com and your licence key arrives by email. Keys you already have keep working, with nothing to change.

#1.3.34

14 September 2026

GoHighLevel connects on the first try. In 1.3.33, connecting with a Private Integration token and Location ID could stop at "Paste the sub-account's Private Integration token." even though the token had been pasted. It now connects the first time.

A changelog in Settings. Settings, Changelog lists what changed in each release, newest first, with the version you have marked.

A clearer message when Codex refuses a model. When your ChatGPT plan does not offer the model a Codex bot is set to, the bot now says so and asks you to pick another, instead of showing Codex's own error. GPT-5.3 Codex Spark, which ChatGPT accounts cannot use, is no longer offered when Codex cannot list its models.

#1.3.33

13 September 2026

Everyone in a channel answers at once. When a message is for several bots, they now work at the same time instead of waiting for each other, and the channel shows who is replying. Team routines do the same: teammates the lead hands work to in one message work together, then report back.

GoHighLevel, built in. Connect your HighLevel sub-account with a Private Integration token and Location ID, or point Pragma at your own HighLevel MCP server. Your bots can then work with contacts, conversations, opportunities and calendars. The token stays in Pragma's encrypted store and never reaches a bot.

A new Connected apps. Apps are grouped by category with search, a Featured row and a list of what you have connected.

Teach a skill without an AssemblyAI key, and on Windows. Recording now works on Windows, with or without a key. After you stop, Pragma can turn your voice into text on your own computer (a one-time 190 MB download you choose to start), or you can type your notes. An AssemblyAI key still gives live captions.

Pick the effort. The model picker now has an Effort row for Claude, Codex and Antigravity, showing only the levels the chosen model supports.

Codex bots start faster. Codex no longer starts your own Codex MCP servers on every turn, the same change Claude got in 1.3.31. Turn on My Codex tools in Settings, Engines if you want them back.

Join Chat. A new link in the sidebar opens the Pragma community on Whop.

Also fixed: a stale duplicate of a bot's reply in channels, bots left busy after deleting a channel or bot mid-reply, two bots on different local models stopping each other's model, and Claude bots in channels re-reading the whole channel every turn.

#1.3.32

13 September 2026

Your message shows the moment you send it. A message used to appear only once the server had sent it back, which in a busy window could take a second. It now shows as you press Enter, and the saved copy takes its place without appearing twice.

Know-how built in. Pragma now comes with its own know-how for design, motion and video work. A bot uses it only when the work calls for it, so it adds nothing to an ordinary message, and what it learns about the designs you study is kept through updates.

#1.3.31

13 September 2026

Channels answer much sooner. Every reply in a channel started a fresh copy of Claude Code, and each copy first loaded the tools from your own Claude Code setup: its MCP servers, plugins and claude.ai connectors, over 200 tools in all, including ones that fail to connect. In a two-bot channel that could take 20 seconds before the first word. Pragma's bots now start Claude Code with Pragma's own tools only, which brings it up in about a second and makes each message cheaper. To let your bots use your own Claude Code tools, turn on My Claude Code tools next to Claude in Settings, Engines.

Bots stop checking their memory before every reply. A bot whose memory was still empty was told its memory file existed but not that it was empty, so a Gemini bot on Antigravity opened the file twice before answering anything, even a greeting. It now hears that there is nothing to read yet, and answers one to four seconds sooner.

Channels are easier to read. Each bot's messages sit under its avatar, name, role and time; the header lists who is in the channel and who is replying; and the typing dots say who is thinking. Status lines no longer look like messages, and a failure reads as an error.

#1.3.30

12 September 2026

Team routines. A routine can now be run by a team instead of one agent. Choose A team under Who does it?, pick two or more agents, make one the lead, and give each its own goal. Each run happens in the routine's own channel: the lead hands out the work by mentioning teammates, they report back, and the lead ends the run with the result. A turn limit and the routine's calendar block keep a run from going on forever; at either one the lead wraps up and says what is left undone. You can watch the channel and write in it while the team works, and the run's receipt shows the team, the turns taken and the result.

Code in a conversation is readable on the light skins. Code blocks were coloured for a dark background whatever the skin, so on Pragma, Atelier and Lagoon their text was near-white on pale grey. They now take light-skin colours on a light skin and dark ones on a dark skin.

#1.3.29

12 September 2026

Connected apps say when a connection has stopped working. A Gmail connection that Google had revoked still looked connected, and a local model asked to check Gmail searched for it again and again until its conversation filled up and it lost track. Now, when an app a bot needs has no working connection, Pragma shows the card to connect it and the bot stops there; the task carries on once it is connected. Connected apps shows an expired connection as expired, with Reconnect.

Local models get 128K. Pragma's own models now start with a 128K context window instead of 64K, in about the same graphics memory, by keeping their context cache at 8 bits. A local model's connected-app searches also come back trimmed to what it needs. On a graphics card that cannot run the smaller cache, the window stays at 64K.

Bring your own model. Settings, Local models now takes any GGUF model from Hugging Face, by its name or a link, or a GGUF file already on this computer. Pragma reads what a model says about itself before anything downloads, turns away one that cannot drive an agent, and says how it would fit this computer. Models you bring appear under On this computer in the model picker beside Pragma's own.

#1.3.28

12 September 2026

Claude Code and Codex run on Pragma's own model. Pragma's own local models worked only with the Qwen engine. Claude Code and Codex now list them under On this computer too. Both were turned away at first: the model's chat template refused any system message after the first one, both engines send such messages, and so every request failed before a word came back. Pragma now starts the model with that one rule relaxed.

A second model, and which one fits your PC. Qwen3.5 4B joins Qwen3.5 9B: a 2.7 GB download for smaller graphics cards. Each model now says whether it fits your graphics card, runs partly in system memory, or is too big for this PC, worked out from this computer's graphics card and memory.

Downloads you can pause, in a folder you choose. A download can be paused and resumed from where it stopped. One that would not fit on its drive is refused before it starts, rather than failing partway. Settings, Local models can move the models to another folder or drive, and new downloads go there too.

Support reports include the local model. Its state and llama.cpp's last lines now go with a support report, so a model that will not load comes with the reason.

#1.3.27

12 September 2026

Connected apps work with a local model. Asked to check Gmail, a local model in the Qwen engine never reached your connected apps. Qwen Code kept their tools behind a search step of its own, which a small model does not take, and loaded them only after its first reply had already gone; so the model reached for a Gmail skill that needed setup of its own, and gave up. Pragma now has Qwen load your connected apps before it starts, hands their tools over with everything else, and tells the model to use them before a skill for the same app.

A local model has room to work. Qwen Code sent 66 tools of its own with every message, about 24,000 tokens, most of them for things Pragma does its own way. Pragma now gives it the 14 a bot works with, so a conversation starts about 20,000 tokens lighter: a model with a 32K window has room for it, and the first reply on Pragma's own model comes sooner. Pragma's own model is also told the window it really has, so a long conversation is summed up before it overflows.

Download Pragma's own model from the model picker. Under the Qwen engine, On this computer lists the models Pragma runs itself, with a Download button and its progress right there, and picking one is a click once it is downloaded. The "too small" message for a local model now points there too.

Your Whop picture follows your account. Pragma asked Whop for your picture only when you signed in, so a picture added later never arrived and your profile kept your initials. It now checks each time Pragma starts, and your profile picture follows it unless you chose a different one in Pragma.

#1.3.26

11 September 2026

Run a model on this computer, with nothing else to install. Local models used to need LM Studio or Ollama installed, running and set up right. Now Settings, Local models downloads Qwen3.5 9B and a copy of llama.cpp, and Pragma runs it itself: on this computer only, with the 64K window the engines need, starting when a bot uses it and freeing the graphics card after 15 minutes without use. Choose it in the Qwen engine as qwen3.5-9b (Pragma). It is Windows only for now, and other engines follow once each has been tested with it.

#1.3.25

11 September 2026

Connect no longer trips over an app you already connected. Clicking Connect in Connected apps before the list of your connections had arrived showed Connect on an app that already had an account. The click was refused with "add an account alias", and the card stayed wrong until the panel was opened again. Connect now waits the moment it takes for that list. If a Connect is still refused because the app already has an account, the card shows the account and offers to add another one under a label. And if the list cannot be loaded at all, the panel says so rather than showing every app as unconnected.

#1.3.24

11 September 2026

Connected apps with your own Composio key open in a second. With a Composio key of your own, the connected list read Composio's whole catalog, over 1,500 apps, before it could say what you had connected: about twenty seconds every time, and it would have stopped working as the catalog grew. It now asks only for the apps you have connected. Checking an app you have not connected no longer fails either.

Fixed on our side, for every version. The "add an account alias" message on Connect, with every app looking disconnected, came from the same catalog reaching our connected-apps service, and was fixed there without an update. Support reports that include an email now get a confirmation with their reference, and the reply comes in the same thread.

#1.3.23

11 September 2026

Your profile fills in from Whop even if you signed in before. 1.3.22 copied your Whop name, email and picture into your profile only at the moment you signed in, so anybody already signed in kept their old details. Now a profile that has never been filled from your Whop account is filled once when Pragma starts, and anything you change after that stays as you left it. The picture comes from your Whop profile when the sign-in itself does not carry one.

Choose what a screenshot or video shows. Support used to capture the Pragma window, and a screenshot could catch Settings on top of it. Now Add a screenshot and Record a video open a picker with previews of the Pragma window, your other windows and each screen, and Settings is fully out of the way before anything is captured.

#1.3.22

11 September 2026

Pragma asks you to sign in. Connected apps belong to the Whop account that signs in, so they stay connected through updates, a reinstall, or a new computer. A copy that was never signed in kept them only as long as its own install lasted. So Pragma now asks for the Whop account you bought it with before anything else. If Whop or our service cannot be reached you can carry on and sign in next time; cancelling the sign-in does not skip it.

Your profile comes from your Whop account. Signing in fills in your name, email and Whop profile picture, and the picture shows in the sidebar. The account lives at the top of Settings, General, Profile now, rather than under Licence.

Connected apps show as connected again. Fixed on our side and already live for every version: Composio started describing an app with no account as null, the status check failed on it, and every app looked disconnected, Gmail included, while the connections themselves were fine.

#1.3.21

10 September 2026

Picking a bigger local model now works. LM Studio loads a model on its first request at its own default context length, 8,192 tokens out of the box, and every Pragma turn sends far more than that before the conversation starts. So choosing a local model that was not already loaded failed every time, even one that can hold 262,144. Pragma now loads an LM Studio model with a 64K window before a turn uses it, or with as much as the model allows if that is less, and the conversation says so while it loads. It only ever touches the model that turn is about to use. If the context will not fit on the graphics card it tries again with the context in system memory, and if LM Studio still cannot load the model, the conversation says so and says what usually causes it.

The explanation arrives whole. An error in the conversation was cut to its first 160 characters, which stopped the too-small-model message at "so a short", before the part that says what to do. Errors now keep up to 600.

#1.3.20

10 September 2026

Support, from inside the app. Settings has a Support section now. Say what went wrong, add a screenshot or a short recording of the Pragma window if it helps show it, and send it; the reply comes by email. Diagnostics go with it unless you untick them: your version, which settings are on (never their keys) and recent log lines, with anything secret removed. A recording is of the Pragma window only, never the rest of your screen, and screenshots, recordings and diagnostics are deleted after 90 days.

A local model that is too small now says so. A model loaded with a small context window used to fail with the raw reply from its host, two layers of JSON deep. Pragma now says how much the model holds, how much the turn needed, and what to do about it. The model picker shows the context window of each loaded LM Studio or Ollama model too, and marks the ones under 64K: the engine sends its own instructions and tools with every message, about 45,000 tokens for Qwen with connected apps on, so a 32K model fails before your message is read.

Connect works again, on every version. Two fixes to the connected-apps service went live earlier the same day, so 1.3.19 has them too: Connect without an account name no longer fails with "invalid JSON body", and a sign-in you started and abandoned no longer blocks the next Connect.

#1.3.19

10 September 2026

Your connected apps can follow you. Until now they belonged to one copy of Pragma. Reinstall it, or set it up on a second computer, and you started again with nothing connected. Settings now has a Your Whop account card in the Licence section. Sign in with the Whop account you bought Pragma with and your connections belong to you instead: every computer you sign in on shares the same set, and signing out of one cuts that computer off from them.

It is a Whop sign-in rather than your licence key on purpose. A key can be shared or leaked, and a key that opened somebody's connected Gmail would turn every leaked key into a leaked inbox.

Connected apps now run on Systo's own service. They used to go through a service run by the open-source project Pragma is built on. Connect your apps once more after this update. The connections you made before were made on that old service and stay there. It is a one-time move: from here on, a reinstall or a new computer only needs you to sign in again.

**Pointing PRAGMA_DATA_DIR at a new folder no longer moves your workspace into it.** The variable exists for testing against a clean workspace, and aimed at a folder that did not exist yet, it used to move the real one there instead. Nothing in Pragma sets it, so this only ever affected someone who set it by hand.

#1.3.18

10 September 2026

LM Studio models show up again when LM Studio is not on port 1234. Pragma looked for LM Studio only at its default address, and LM Studio's port is a setting: anyone who had ever changed it, or whose 1234 was once taken, saw "No local models found" with the server running and models downloaded. Pragma now reads the port from LM Studio's own settings, and the agent you pick is started against the same address the picker found, so a model cannot be listed from one port and then run against another.

Qwen stops collecting stale provider entries. When a local host moved, the Qwen setup kept the old entry and added a second one with the same model name, and Qwen had nothing but order to choose between them. It now updates the entry in place.

LM Studio's own server still has to be running for any of this to work. If it is set not to start with the app, turn that on in LM Studio's server settings or run lms server start.

#1.3.17

10 September 2026

The new app icon. The speech bubble that the site, the storefront and the app title bar already used, now on the installer, the taskbar and the desktop shortcut. 1.3.16 and earlier still carry the older rounded face, because the icon was redrawn after that build was cut.

Nothing else in the app changed. Two website fixes shipped alongside it: the banner no longer says Pragma is free while we launch, which stopped being true the day the checkout went up, and update downloads are differential again, so this release is the first that arrives as a delta rather than the whole hundred megabytes.

#1.3.16

9 September 2026

Settings can now tell you what the last update check actually did. It used to say "you're on the latest version we know of" whether a check had succeeded or had never run at all, because the status resets to idle the moment a check ends. There is a last-checked line with the outcome now, and a button that reveals the update log, since asking somebody to send their log only helps if they can find it.

A failed check says one sentence instead of nine hundred characters. electron-updater reports a failure as the whole HTTP exchange, status line, body and every response header, which on this server includes the entire Content-Security-Policy. "This copy of Pragma has no licence key" is the part that matters.

The update log now complains when it cannot be written. It swallowed write failures silently before. That is the wrong behaviour for the one file that would show an update problem, on a product whose updater was inert for fifteen releases without anyone noticing.

Nothing else in the app changed. The documentation site, the shorter-lived download links and the receipt page's Whop sign-in are all on the website and were live before this build.

#1.3.15

9 September 2026

The updater actually runs. Pragma checks fifteen seconds after launch and then hourly, against its own feed, which verifies your licence before it serves anything. A card appears in the bottom-left corner when there is something to take. Nothing installs without a click. See Updates.

A licence panel in Settings. Paste the key from your purchase once. It is stored with the operating system's own encryption, never in plain text, and it is what the updater sends on every check.

The updater's fallback is no longer an open door. The configuration baked into the app still named the public releases repository, so if the runtime override had ever failed, the app would have quietly gone back to an ungated feed. Both layers name the licensed feed now, and the failure mode is a refused update rather than a silent bypass.

#1.3.14

8 September 2026

The app is unchanged from 1.3.13. This release exists so the installer and the website carry the same version number.

#1.3.13

7 September 2026

Housekeeping. Nothing changes what the app does.

The iPhone app is gone from the source tree. It was written, tested, and never distributed, so there was no way to install it, and every hour it stayed cost a CI job on a macOS runner, a Swift package to keep compiling, and a fixture generator to keep in step with a server contract, all for a client nobody could run. About sixteen thousand lines removed, recoverable from history if a native client is ever worth building again.

The pragma://pair link went with it. A custom scheme with no app to claim it is a QR code that scans to an error.

#1.3.12

7 September 2026

A review pass over the companion, and what it found.

• An agent's profile picture was a broken image on your phone. Every companion route needs its token and an image tag carries no headers, so a custom avatar came back a 401. The coloured initial is drawn first now, with the picture layered over once it arrives.
• Opening a chat did not land on the newest message. The jump to the bottom happened, then images loaded underneath and pushed it back up. A late image now returns you to the bottom, unless you had already scrolled back.
• Older messages are reachable. The phone fetched the newest eighty and stopped, so a long conversation looked like it began there.
• An answered approval could sit with both buttons dead.
• A hidden agent came back into the list, and re-pairing could open a second event stream that folded every message in twice.

#1.3.11

7 September 2026

The companion told your phone an address it could never reach. On Windows with WSL or Hyper-V installed, the panel printed the virtual switch: a network with one machine on it and no way in. Virtual adapters are recognised now, and the routing table gets the final say, since asking the kernel which address it would use to reach the internet is the same question as which network this computer is really on.

A QR code to open it. Typing an IP and a port on a phone keyboard is the worst part of the setup. The code carries no pairing secret, so it can be scanned before pairing starts, which is the order people actually do it in.

Images show up on the phone. A generated picture used to appear as a chip reading "Screen frame". Tap to open it full size.

An agent no longer promises lettering it cannot draw. When image generation points at a generator on your own machine, the tool says so, and says that words come back as gibberish and to set type separately.

#1.3.8

7 September 2026

A generated image is a picture again, and you can open it. The picture was always fine; the wrapper was not. Two code paths wrote a whole data URL into a field holding bare base64, so the chat wrapped a data URL in another data URL and the image endpoint decoded the prefix along with the payload. Images already in your transcripts come back too.

Click one and it opens. Full window, click again for actual size and pan, Save writes a copy under the filename the agent gave it.

A generated image is no longer pruned. The frame keeper holds the newest few screenshots from computer use and drops the rest, which is right for a live screen and is deletion for a poster you asked for. They travelled as the same kind of message with nothing to tell them apart.

#1.3.7

7 September 2026

Your phone can open Pragma through its browser, with no app to install. Settings, Companion, start pairing, then scan the code. See The web companion.

#Earlier

1.0.1 through 1.3.6 were released across 6 and 7 September 2026: the rename from the upstream fork, the first Pragma-branded Windows installer, and the run of fixes that followed it. Their notes are on the releases page.