MCP servers give Codex CLI access to external tools and data over the Model Context Protocol that Claude Code uses too. Codex stores them in TOML, and the CLI, the IDE extension and the ChatGPT desktop app all share that config. Everything below is from the official Codex MCP docs.
1. Find the config file
Codex reads ~/.codex/config.toml for user-wide servers. A project can add .codex/config.toml, but it applies only to trusted projects. Create the file if it does not exist.
2. Add a filesystem server
Add one table per server, named [mcp_servers.<name>]:
[mcp_servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects"]
Here command is required, and args, env, env_vars, cwd, enabled and required are optional. Timeouts default to startup_timeout_sec = 10 and tool_timeout_sec = 60.
3. Or let the CLI write it
This adds the same STDIO server without editing TOML by hand:
codex mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects
Pass --env KEY=VALUE before the -- to forward an environment variable.
4. Add a remote HTTP server
Streamable HTTP servers use a url instead of a command:
[mcp_servers.example]
url = "https://mcp.example.com/mcp"
bearer_token_env_var = "EXAMPLE_TOKEN"
For OAuth servers, run codex mcp login example once to authenticate.
5. Launch and confirm it loaded
List servers from your shell, or type /mcp inside the Codex TUI:
codex mcp list
A healthy server appears with its tools, and Codex can call them during a task. The Claude Code analog is install an MCP server in Claude Code.
Verify it worked
- Server missing from
codex mcp listCheck the table name is exactly[mcp_servers.name], then restart Codex. spawn npx ENOENTNode is not on PATH; install Node or givecommandan absolute path.- Remote server returns 401 The token is missing; set the variable named in
bearer_token_env_var, or runcodex mcp login.
Source: Codex MCP docs.