Connecting Backends

Step-by-step setup guides for connecting Ghost OS and agent-browser backends to ADE, plus computer use policy reference.

Connecting a Backend

Ghost OS Setup

1. Install Ghost OS

   Download and install the Ghost OS MCP server for macOS from ghostos.dev. Start the server — it listens on https://ghost-os.local/mcp by default.

2. Add to ADE configuration

   Open .ade/local.secret.yaml in your project root (create it if it does not exist) and add the Ghost OS entry:

   externalMcp:
     - name: ghost-os
       transport: http
       url: https://ghost-os.local/mcp
       autoStart: true
   
3. Verify connection

   Open Settings → Computer Use. The Ghost OS card should show a green Connected badge with the capability matrix populated. If it shows an error, confirm the Ghost OS server is running and the URL is reachable.

4. Set as preferred backend

   In Settings → Computer Use → Preferred Backend, select Ghost OS. ADE will use Ghost OS for all computer-use requests until you change this setting.

agent-browser Setup

1. Ensure Node.js 18+ is installed

   Run node --version to confirm. If needed, install via nodejs.org or brew install node.

2. Add to ADE configuration

   externalMcp:
     - name: agent-browser
       transport: stdio
       command: npx
       args: ["-y", "@agent-browser/cli"]
       autoStart: true
   
3. Verify connection

   Open Settings → Computer Use. The agent-browser card should show Ready. On first use, ADE will download the @agent-browser/cli package automatically.

---

Computer Use Policies

Each chat session and mission has a computer-use policy header that controls when the agent is allowed to invoke computer-use tools. Set the policy at session start or change it mid-session from the chat header bar.

Policy	Behavior
CU Off	Computer use tools are never available in this session. Agents cannot invoke them even if they try.
CU Auto	ADE decides when computer use is appropriate based on the task. The agent does not choose directly.
CU On	Computer use tools are always available. The agent decides when to invoke them.
Fallback	The agent attempts the task without computer use. If it cannot complete the task, it falls back to CU.
Proof	Computer use is available, but every action must produce a proof artifact. The session blocks if proof is missing.

Use Proof mode for any automation that will be shared with stakeholders or linked to a Linear issue. The collected artifacts provide an auditable record of exactly what the agent did.