Start here

A tour of the window

Six parts of the interface, and what each one is for.

A fresh Pragma agent, ready for its first message
A new agent on first run. Each conversation belongs to a real process, not a simulated contact.

The sidebar

Your agents, in the order you arrange them. Search at the top looks through both agent names and message history, so an error string or an issue number finds the conversation it appeared in.

Below the roster: Teach a skill, Tasks and routines, Connected apps, and your profile with app settings.

The conversation

The work log. Replies, questions, approval requests, tool progress, attachments and failures all stay in the conversation that produced them, in order.

Messages can be pinned, copied, reacted to, retried and read aloud. Editing an earlier message starts a new branch rather than rewriting history, and Pragma deliberately does not resume the old provider session afterwards, so the next turn follows the branch you can see rather than one you cannot.

The actions available on a message
Pin a decision you will need again. Retry when a connection dropped rather than when an answer was wrong.

Tasks

The Task button starts a separate unit of work without disturbing the main conversation. A task has its own transcript, its own provider session, its own approvals and its own cost.

Use one when the work has a clear beginning and a result you want to point at later. Scheduled runs create tasks on their own.

The model picker

Top of the conversation. It groups engines by provider, marks which are installed and signed in on this machine, and lets each agent use something different.

The model picker, showing cloud and local engines
Models are chosen per agent, so a team can mix engines without touching a global setting.

Changing the model affects the next turn. It does not rewrite what has already been said. The Effort row under the models sets how hard the chosen model thinks, where the model allows it.

The computer panel and the inspector

Computer decides where visual work happens: nowhere, this computer, an isolated local VM, or a hosted desktop. It is a per-agent choice and it is off by default. See Giving an agent a computer.

Inspector is for when the conversation is not enough detail: tool calls, reconnects, raw engine output, long-running work.

Connected apps and routines

Connected apps turns an authorised account into tools an agent can use; GoHighLevel is built in. Tasks and routines turns a prompt that worked into one that runs every weekday morning.

Settings

The gear by your name. Engines and their paths, API keys for optional services, appearance, and your licence.

Pragma app settings
Credentials entered here are validated before they are saved, and stored with the operating system rather than in a config file.