# codex-as-mcp

[中文版](./README.zh-CN.md)

`codex-as-mcp` is a small **Model Context Protocol (MCP)** server that lets MCP clients (Claude Code, Cursor, etc.) delegate work to the **Codex CLI**.

It exposes two tools that run Codex in the server's current working directory:
- `spawn_agent(prompt: str)`
- `spawn_agents_parallel(agents: list[dict])`

Under the hood, each agent runs something like:
`codex exec --cd <server cwd> --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox -`.

The prompt is sent through stdin and then closed. This avoids command-line quoting/length issues and prevents Codex from accidentally reading from the MCP server's JSON-RPC stdin pipe.

Note: `--dangerously-bypass-approvals-and-sandbox` disables sandboxing and confirmation prompts. Use this server only in repos you trust.

## Use it in Claude Code

There are two tools in codex-as-mcp
![tools](assets/tools.png)

You can spawn parallel codex subagents using prompt.
![alt text](assets/claude.png)

Here's a sample Codex session delegating two tasks in parallel.
![Codex use case](assets/codex.png)

## Quick start

### 1. Install Codex CLI

**Requires Codex CLI >= 0.46.0**

```bash
npm install -g @openai/codex@latest
codex login

# Verify installation
codex --version
```

Make sure Codex CLI can run non-interactively on your machine (provider + credentials in `~/.codex/config.toml`, or via the provider-specific env var it references).

#### Example: third-party provider + `env_key`

If you're using a third-party provider, configure it in Codex `config.toml` and point `model_provider` at it. When a provider uses `env_key`, Codex CLI expects that env var to be present when it runs.

Example:
```toml
model_provider = "custom_provider"

[model_providers.custom_provider]
name = "custom_provider"
base_url = "https://..."
wire_api = "responses"
env_key = "PROVIDER_API_KEY"
show_raw_agent_reasoning = true
```

When using `codex-as-mcp`, make sure the MCP server process has that env var set, so it can pass it through to the spawned `codex` process. The env var name **must match** the `env_key` value above (here: `PROVIDER_API_KEY`).

**Option A (recommended): set env in your MCP client config (if supported)**
```json
{
  "mcpServers": {
    "codex-subagent": {
      "type": "stdio",
      "command": "uvx",
      "args": ["codex-as-mcp@latest"],
      "env": {
        "PROVIDER_API_KEY": "KEY_VALUE"
      }
    }
  }
}
```

**Option B: pass env via server args**
```bash
uvx codex-as-mcp@latest --env PROVIDER_API_KEY=KEY_VALUE
```

**Option C: add via Codex CLI (`codex mcp add`)**
```bash
codex mcp add codex-subagent --env PROVIDER_API_KEY=KEY_VALUE -- uvx codex-as-mcp@latest
```

Security note: passing secrets via command-line args may be visible via process lists on your machine; prefer option A when possible.

### 2. Configure MCP

Add to your `.mcp.json`:
```json
{
  "mcpServers": {
    "codex-subagent": {
      "type": "stdio",
      "command": "uvx",
      "args": ["codex-as-mcp@latest"]
    }
  }
}
```

Or use Claude Desktop commands:
```bash
claude mcp add codex-subagent -- uvx codex-as-mcp@latest
```

If you're configuring Codex CLI directly (for example `~/.config/codex/config.toml`), add:
```toml
[mcp_servers.subagents]
transport = "stdio"
command = "uvx"
args = ["codex-as-mcp@latest"]
# Increase if you see ~60s tool-call timeouts when running longer Codex tasks.
# tool_timeout_sec = 600
```

## Tools

- `spawn_agent(prompt: str)` – Spawns an autonomous Codex subagent using the server's working directory and returns the agent's final message.
- `spawn_agents_parallel(agents: list[dict])` – Spawns multiple Codex subagents in parallel; each item must include a `prompt` key and results include either an `output` or an `error` per agent.

## Maintainer live smoke test

The default CI uses unit tests and a fake subprocess path so it can run safely without credentials. Maintainers can additionally run the manually triggered **Live Codex smoke** GitHub Actions workflow with a real provider by setting repository secret `CODEX_LIVE_API_KEY` and providing a Responses-compatible `base_url`/`model` in the workflow inputs.

Note: current Codex CLI rejects `wire_api = "chat"`; chat-completions-only providers such as direct DeepSeek API need either native Codex support for that wire API or a Responses-compatible proxy.

## Troubleshooting

### `spawn_agent` times out after ~60s

If you see an error like:
```text
tool call failed for `subagents/spawn_agent`
timed out awaiting tools/call after 60s
deadline has elapsed
```

This is typically a client-side MCP tool-call timeout. `spawn_agent` does not return until the spawned `codex exec` process finishes, which can take longer than 60 seconds.

Fix: increase the tool-call timeout in your MCP client.

#### Codex CLI

In your Codex config (`~/.codex/config.toml` or `~/.config/codex/config.toml`), set a higher `tool_timeout_sec` for the MCP server:
```toml
[mcp_servers.subagents]
transport = "stdio"
command = "uvx"
args = ["codex-as-mcp@latest"]
tool_timeout_sec = 600
```

#### MCP Inspector / `mcp dev`

If you're testing locally with the MCP Inspector, increase request timeouts (or run `./test.sh`, which exports these):
```bash
export MCP_SERVER_REQUEST_TIMEOUT=300000
export MCP_REQUEST_TIMEOUT_RESET_ON_PROGRESS=true
export MCP_REQUEST_MAX_TOTAL_TIMEOUT=28800000
```