## Using Codex as an MCP Server The Codex CLI can also be run as an MCP _server_ via `codex mcp-server`. For example, you can use `codex mcp-server` to make Codex available as a tool inside of a multi-agent framework like the OpenAI [Agents SDK](https://platform.openai.com/docs/guides/agents). Use `codex mcp` separately to add/list/get/remove MCP server launchers in your configuration. ### Codex MCP Server Quickstart You can launch a Codex MCP server with the [Model Context Protocol Inspector](https://modelcontextprotocol.io/legacy/tools/inspector): ```bash npx @modelcontextprotocol/inspector codex mcp-server ``` Send a `tools/list` request and you will see that there are two tools available: **`codex`** - Run a Codex session. Accepts configuration parameters matching the Codex Config struct. The `codex` tool takes the following properties: | Property | Type | Description | | ----------------------- | ------ | ------------------------------------------------------------------------------------------------------------------------------------------------------ | | **`prompt`** (required) | string | The initial user prompt to start the Codex conversation. | | `approval-policy` | string | Approval policy for shell commands generated by the model: `untrusted`, `on-failure`, `on-request`, `never`. | | `base-instructions` | string | The set of instructions to use instead of the default ones. | | `config` | object | Individual [config settings](https://github.com/openai/codex/blob/main/docs/config.md#config) that will override what is in `$CODEX_HOME/config.toml`. | | `cwd` | string | Working directory for the session. If relative, resolved against the server process's current directory. | | `model` | string | Optional override for the model name (e.g. `o3`, `o4-mini`). | | `profile` | string | Configuration profile from `config.toml` to specify default options. | | `sandbox` | string | Sandbox mode: `read-only`, `workspace-write`, or `danger-full-access`. | The `tools/call` response for `codex` includes the thread id in `structuredContent` so the client can resume the session later: ```json { "content": [{ "type": "text", "text": "..." }], "structuredContent": { "threadId": "..." } } ``` While the tool is running, `codex/event` notifications include `_meta.threadId` so clients can correlate events to a conversation: ```json { "_meta": { "requestId": 1, "threadId": "..." }, "id": "evt-...", "msg": { "type": "..." } } ``` **`codex-reply`** - Continue a Codex session by providing the thread id and prompt. The `codex-reply` tool takes the following properties: | Property | Type | Description | | ------------------------- | ------ | -------------------------------------------------------- | | **`prompt`** (required) | string | The next user prompt to continue the Codex conversation. | | **`threadId`** (required) | string | The id of the conversation to continue. | ### Trying it Out > [!TIP] > Codex often takes a few minutes to run. To accommodate this, adjust the MCP inspector's Request and Total timeouts to 600000ms (10 minutes) under ⛭ Configuration. Use the MCP inspector and `codex mcp-server` to build a simple tic-tac-toe game with the following settings: **approval-policy:** never **prompt:** Implement a simple tic-tac-toe game with HTML, JavaScript, and CSS. Write the game in a single file called index.html. **sandbox:** workspace-write Click "Run Tool" and you should see a list of events emitted from the Codex MCP server as it builds the game.