OpenAI Codex (the Codex CLI) consumes RAC on the same two surfaces it already supports natively — an AGENTS.md it reads for project instructions, and MCP servers it connects to — so there is no Codex-specific RAC code. A stranger can reproduce this from the file alone.

pip install rac-core   # the `rac` CLI and the `lore` MCP server

A repository with a RAC corpus under decisions/ (run decided quickstart, or use this repository's own decisions/).

Codex reads AGENTS.md for project instructions. RAC generates one from your recorded decisions:

decided export decisions/ --agent-rules

Writes AGENTS.md at the repo root, in a managed block (your own content is preserved). Codex discovers it from the project root; re-run the export when decisions change (decided export decisions/ --agent-rules --check fails CI on drift).

Either use the CLI:

codex mcp add lore -- decided-mcp --root .

…or add a [mcp_servers.lore] table to Codex's config.toml (a sample is in config.example.toml):

[mcp_servers.lore]
command = "rac"
args = ["mcp", "--root", "."]

• Global config (most reliable): ~/.codex/config.toml — use an absolute --root path there.
• Project config: .codex/config.toml in the repo root is honoured for trusted projects; some Codex surfaces only load the global file, so if the server does not appear, fall back to ~/.codex/config.toml.

This exposes the five read-only lore tools (get_summary, search_artifacts, get_artifact, get_related, find_decisions). The server re-reads the corpus on every call and never writes to the repo.

RAC supplies context and enforces after the edit (ADR-067) — it does not intercept Codex's loop. Whatever Codex writes is checked by decided validate and decided relationships --validate (and the GitHub Action / pre-merge gate) the same as any other contributor; the trust boundary is human PR review and CI. The per-edit pre-edit veto is Claude-Code-specific (see examples/claude-code/); with Codex you rely on the CI gate, which keeps Codex's loop untouched and adds no latency.

Run the bundled grounding demo — same task twice, once unconnected and once with lore connected — and watch the connected run respect a recorded decision the unconnected run violates: examples/guide/.