codex

Driving CEL from OpenAI Codex CLI

This page shows how to drive CEL from the OpenAI Codex CLI.

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

Purpose

Codex CLI is OpenAI's terminal-resident agent. It calls tools, loops, and stops on its own, which is exactly the external-agent pattern CEL is designed for. In this setup:

• Codex owns the loop and tool-calling.
• CEL provides cel_see, cel_act, cel_perceive, cel_think via MCP stdio.
• You normally only need cel_see + cel_act.

Setup

1. Build CEL

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

2. Register CEL with Codex

Codex exposes MCP servers via a config file in the user's home directory.

TODO: verify with latest Codex CLI docs at https://platform.openai.com/docs/codex (or the Codex GitHub repo README) — the exact path and key name have changed between Codex releases. As of the most recent public guidance the config lives under ~/.codex/ and MCP servers follow the standard mcpServers schema shared across MCP clients.

The MCP server block itself is stable:

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

Or via the CLI entry point:

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

3. Verify Codex sees CEL

Start Codex and list tools. The exact command varies by version; check codex --help or codex mcp list in your build.

TODO: verify with upstream docs — Codex version in April 2026 should support listing attached MCP servers. If your version does not, run the MCP Inspector against CEL directly to confirm the server is healthy:

npx @modelcontextprotocol/inspector node /path/to/cellar/mcp-server/dist/index.js

4. Grant macOS Accessibility permission

As with every CEL client, the first cel_act call that performs input requires Accessibility permission for the host process. Grant it in System Settings → Privacy & Security → Accessibility and restart Codex.