With an agent

Your AI agent uses Daimon through its MCP tools: it looks at the screen, reads the interface, acts on your behalf and shows you what it is doing. You decide, the agent executes, and the ceiling you set bounds everything it can do.

You talk to your agent in natural language. Each page of this guide starts from what you ask it, then says what it does, what you get, and what happens when a call is refused.

Compatibility

Daimon does not target one particular agent: any MCP client that can launch a server over stdio can use it. Agents backed by frontier models are the most likely to chain perception and action reliably. The list of clients Daimon knows how to configure is in AI clients.

Working at your best

Unlike a tool tied to a project, Daimon has no working folder: it acts on your machine, wherever the agent's session is open. What matters:

• Daimon is registered in your agent and the agent has been restarted since (see below);
• the system permissions are granted: Permissions;
• the ceiling matches the task: at L0 (the default setting), the agent can look but not act. See Acting;
• the Daimon panel is within reach, and Présence à l'écran (On-screen presence) is turned on if you want to see what the agent is doing: see Showing what is happening.

Then talk to it: "Check that Daimon is ready" or "Look at my screen and tell me which window is in the foreground".

Connecting an agent

Three routes, which do the same job, detailed in AI clients:

• the setup flow, then the panel, with one switch per client;
• a terminal: Daimon install registers Daimon in every detected client;
• for an unknown client, a stdio declaration whose command is the Daimon binary and whose argument is serve.

After registering, restart your agent so that it sees the tools.

You stay in control

• The ceiling. Each Mains (Hands) action has a level (L1 to L4). The agent can read the ceiling with main_ceiling, but never raise it.
• Confirmation. Up to L3, an action that Daimon judges irreversible stops on a confirmation dialog, which declines by itself after 30 seconds.
• L4 is your decision. It requires your explicit consent, from the panel or the terminal. No tool lets the agent engage it.
• Your exclusions. The applications and areas you protect are censored or refused: Ceiling and exclusions.
• The audit. Every action is logged with the intent the agent declared.

When a call is refused

The refusal always says why, in a form your agent understands directly:

Form	Meaning
{"status": "refused", "reason": "…"}	A Mains action is refused: ceiling, guard, confirmation not obtained
{"status": "error", "reason": "…"}	A Mains action failed to run
{"error": "<code>", "message": "…"}	A preparation or perception error: invalid argument, missing permission, protected application

Each journey page lists the actual refusals of its tools. When in doubt, the agent calls daimon_doctor.

The journeys

Page	For
Getting started	Check that Daimon is ready before the first task
Looking at the screen	Capture a screen, locate a text, convert coordinates
Reading an interface	Get a window's elements without going through the image
Acting	Click, type, press, drag, under the ceiling you set
Showing what is happening	See on screen what the agent is doing
Delegating to a sub-agent	Hand a sequence of actions to another agent
When it doesn't work	Find the cause before concluding something is broken

Reference

Every tool, its parameters and its level: Tool reference.