DocsUse Mecum
Workspace
Configure a worker, give it a task and follow its conversation. Messages record what was said; tool events record what the agent requested and what the runtime reported.
Configure a worker
A worker combines a name, role, instructions and a selected agent connection. It does not represent a separate Mac or an already acquired desktop session.
- Open the team workspace and choose New Worker.
- Give it a name. Role and instructions are optional and tell the model its responsibility.
- Choose a provider and model now, or later from the composer.
- Start the conversation with a specific task and a result you can check.
Mecum offers Mac Access at first launch and from Settings. Desktop work requires the relevant grants; text conversations do not.
Claude Code and Codex run their own agent loop and call Mecum’s tools through the local MCP bridge. API model connections use Mecum’s tool loop when the model supports tools. External MCP clients, Claude Code and Codex first, are documented in Models & MCP.
Follow a conversation
Mecum saves a submitted message before starting a worker turn. If saving fails, the message returns to the draft. During the turn, response output and tool activity arrive as separate events.
A worker host does not start a second turn while one is active. Finish or stop the current turn before starting again. Adding workers does not add simultaneous desktop capacity.
Read replies and events
| Entry | Meaning |
|---|---|
| Worker reply | The model’s explanation. Compare a claim of success with the action evidence. |
| Tool activity | The requested operation and reported outcome. Delivered input is not always a confirmed effect. |
| Partial output | Text received before completion. An interrupted stream is not a finished reply. |
| Error or stop | The point where work ended. Check any actions already delivered before retrying. |
For uncertain effects, follow the action evidence guide before sending the same request again.
Stop and inspect
Stop Response, the stop button in the composer or the Team menu, prevents new tool requests, interrupts the worker’s process and waits for the active tool operation to settle. The interrupted turn is marked Stopped and is not retried.
Release Computer, the button beside the composer while the worker holds the computer or the Team menu, gives the desktop seat back: the window returns to your display and an application the worker launched is quit. The conversation, the worker and a running turn stay; the next tool call that needs the computer waits for the seat again.
After a turn, a worker keeps the seat for 30 seconds in case of a follow-up, then releases it. It releases at once when another worker is waiting.
Neither undoes an edit or a submission already delivered. Read the last events and inspect the app before starting another task.
History, context and sessions
Workers, conversations, messages and events are stored locally in Workspace.store. The provider’s session ID is saved with the conversation: when the same provider answers next, Mecum asks it to resume that session; a different provider starts a new one. Whether the resume succeeds depends on the provider still holding the session. A failed resume is reported, never replaced by a silent new conversation.
The context control beside the composer shows how full the model’s context is. Compact context summarizes the conversation so far and keeps going; Mecum compacts on its own when a turn ends above 90%. Start fresh context forgets everything before now; the chat stays.
Saved history cannot restore a window target or desktop access: after a restart, old seat and window references are invalid. If Mecum exits without shutting down cleanly, the next launch marks the interrupted turn as failed, runs nothing again and does not restore window positions.
Read Memory & Brain for app knowledge, or Models & MCP to troubleshoot the connection.
Checked against Mecum app source at main 524eb7f on October 1, 2026. UI labels and paths have not yet been checked against the signed release build.