Codex reads the canonical AGENTS.md natively and supports SKILL.md
skills natively — init installs the same skill body Claude gets, at Codex's
paths. Register the MCP server once globally, or per repo.
~/.codex/config.toml:
[mcp_servers.loopback]
command = "node"
args = ["/ABS/PATH/loopback/dist/index.js"]
# or zero-install: command = "npx" / args = ["-y", "loopback-mcp-server"]HTTP variant — point at the long-running central instance instead:
[mcp_servers.loopback]
url = "http://127.0.0.1:7077/mcp"Skill, machine-wide (optional): copy skills/loopback/SKILL.md from this
repo to ~/.agents/skills/loopback/SKILL.md to make the skill available in
every project without running init.
npx loopback-mcp-server init --project <slug> --writewrites for Codex: the AGENTS.md queue section (read natively), the skill at
.agents/skills/loopback/SKILL.md, and a project-scoped .codex/config.toml
with the MCP entry. Codex loads project-scoped config only after you trust
the repo in Codex — if you skip trust, use the global config above instead
(init prints that block too).
Say "work the feedback queue" — AGENTS.md and the skill both carry the
loop. Explicit trigger: $loopback or "Work the Loopback queue for project
<slug>."