Skip to main content
OpenAI Codex extends through MCP servers declared in ~/.codex/config.toml and AGENTS.md instructions. Wire up the MCP server with Smithers:
That registers the MCP server in Codex’s config. Restart Codex and it can drive workflows.

​
Model tiers

Smithers pins each Codex worker to the tier that matches its role: Use the exact tier IDs rather than the floating gpt-5.6 alias. As of registry v7 substantial implementation defaults to Claude Opus 5, which outbenchmarks Sol on agentic coding at lower output cost; Codex holds the checking seats: Sol for final review and second opinions, Terra for validation and tool-heavy checking, Luna for trivial, minimal-risk work and research gathering only. Codex tiers never hold the orchestration or gating seat: scope, direction, progress, and done-or-not decisions run on Claude (claude-opus-5 by default, claude-fable-5 for planning). When usable Codex authentication is present, generated Smithers workflows put these Codex workers first in the checking pools (review and validation) and directly behind Claude Opus 5 in the implementation pool. Other adapters remain later runtime fallbacks and run only when earlier attempts are unavailable or fail. See SOTA role defaults.

​
Smithers docs in Codex

Smithers does not currently auto-install a Codex skill, and bunx smthrs skills add has no supported --agent codex target filter. For Codex, keep the repo-level AGENTS.md guidance below and use the packaged docs commands when Codex needs the full API:

​
Register the MCP server

mcp add writes the entry for you. The native config is a TOML table in ~/.codex/config.toml:
Or use the Codex CLI:
Codex launches the server at session start and the semantic tools (list_workflows, run_workflow, watch_run, resolve_approval, …) appear in the /mcp list. Full reference: MCP Server. Codex CLI and MCP launches detect CODEX_THREAD_ID (ahead of inherited Claude markers) and persist Codex harness/session attribution. Use the explicit --started-by-* flags or MCP startedBy object when providing additional caller-known provenance; attribution is not authentication.

​
Standing instructions

Codex reads AGENTS.md from the repo root down to the working directory. Add a section so Codex knows to reach for Smithers:
This same AGENTS.md is read by Cursor, Copilot, Hermes, and Pi; one file covers several agents.

​
Smithers runs Codex too

In the other direction, Smithers spawns Codex as a worker via CodexAgent (codex exec with a streamed JSON protocol and native thread hijack):
See CLI Agents → Codex.