Skip to content

README.md

Latest commit

 

History

History
65 lines (49 loc) · 2.91 KB

README.md

File metadata and controls

65 lines (49 loc) · 2.91 KB
Raw

Using Chief from Codex

The same two pieces as Claude Code, doing the same jobs:

  • The MCP server is the capability surface — the same 30 tools, over the same streamable-HTTP endpoint. Codex and Claude Code connect to one server and one database; a workflow planned from one is reportable from the other, and harness on each step says which agent the work was planned for.
  • SKILL.md is the protocol — one file, shared with Claude Code deliberately, because the protocol is a property of Chief, not of the client: plan first, summarise every update, propose an amendment rather than improvising, wait for the human, read what a person wrote back.

1. Run Chief

chief --port 8080          # api, ui and the MCP endpoint, one process

HTTP rather than stdio matters more for Codex than the convenience: a stdio server is spawned as a child process per client, and a second process on the same SQLite file is exactly the corruption the store's lock exists to prevent. One long-running Chief, every client a network caller.

2. Register the server

codex mcp add chief --url http://127.0.0.1:8080/mcp/

This writes [mcp_servers.chief] into ~/.codex/config.toml, which the CLI, the IDE extension and the desktop app all share. Verify with codex mcp list, or ask Codex to call list_workflows.

3. Install the skill

mkdir -p ~/.codex/skills/chief
ln -s "$PWD/integrations/SKILL.md" ~/.codex/skills/chief/SKILL.md

A symlink for the same reason as on the Claude side: the protocol the skill describes is enforced by the code next to it, and the two drifting apart is the failure worth avoiding.

Codex has no slash-trigger; it surfaces skills by name and description. Asking to "track this in Chief" is what brings it in — the same opt-in-per-task rule as everywhere else.

Headless runs

Codex's sandbox counts the MCP HTTP call as network access. In an interactive session that is a prompt to approve; under codex exec, where nobody can answer, a restrictive sandbox auto-denies it and every Chief tool fails with "user cancelled MCP tool call" — which reads as a Chief problem and is not one. Headless sessions that need Chief must run with a sandbox that allows network (e.g. -s danger-full-access inside an environment that is already contained); interactive sessions need nothing special.

What Codex cannot do

Exactly what Claude cannot, for exactly the reasons in integrations/claude-code/README.md and MCP-SURFACE.md: the approval policy, the audit log, artifact comments, review notes, re-filing and renaming are REST-only, and the four decision tools (approve_workflow, approve_amendment, reject_amendment, resolve_checkpoint) are human decisions to be called only on an explicit instruction in the turn. Chief records which transport each decision arrived on either way.