claude-code

Driving CEL from Claude Code

This page shows how to drive CEL from Claude Code.

Read docs/adapters-cel-agents.md first if you haven't.

Purpose

Claude Code is a strong default client for CEL because:

• it is MCP-native (stdio transport, same boundary CEL exposes)
• it already owns tool calling, retries, stop conditions, and user approval
• you do not need to bring your own LLM or planner wiring

In this setup Claude Code owns the loop, and CEL is the trust/execution server providing cel_see, cel_act, and cel_perceive. cel_act returns receipts; Claude Code should use those receipts plus post-action evidence in the final answer. cel_think is available but should usually stay unused here — Claude Code already plans.

Setup

1. Build CEL

cd /path/to/cellar
pnpm install && pnpm -r build

2. Register CEL as an MCP server

Claude Code supports two config locations:

• User-wide: ~/.claude/settings.json
• Project-scoped: ./.mcp.json in the repo root

Project-scoped is recommended for CEL because permissions and registry state are machine-local.

Add CEL to the MCP section:

{
  "mcpServers": {
    "cel": {
      "command": "node",
      "args": ["/absolute/path/to/cellar/mcp-server/dist/index.js"]
    }
  }
}

Equivalent CLI form (if the cellar binary is on your PATH):

{
  "mcpServers": {
    "cel": {
      "command": "cellar",
      "args": ["mcp"]
    }
  }
}

3. Start Claude Code

Claude Code picks up MCP servers at start. In a new session, run:

/mcp

You should see cel listed with four tools: cel_see, cel_act, cel_perceive, cel_think.

TODO: screenshot — Claude Code /mcp output showing CEL registered.

4. Grant Accessibility permission

The first cel_act call will fail until macOS grants Accessibility permission to whichever process is launching Node (Terminal, iTerm, the Claude Code host, etc.). Grant it in System Settings → Privacy & Security → Accessibility, then restart Claude Code.