OpenAI Codex CLI integration for CodraGraph. Enriches every Codex tool call with graph-aware context (callers, impact, process participation) and FeatureCluster packs for product areas like Settings/Auth/AI, so Codex doesn't blindly grep its way through the codebase.

Install

# 1) Install the codragraph CLI
npm install -g @codragraph/cli
# or: bun add -g @codragraph/cli --trust
codragraph setup

# 2) Install this integration globally
npm install -g @codragraph/codex
# or: bun add -g @codragraph/codex --trust

# 3) Wire it into Codex. Merges the bundled hooks into ~/.codex/config.json
#    (substituting the absolute install path so Codex finds the hook script)
#    and registers an mcpServer entry whose launcher is platform-correct:
#    `codragraph mcp` on macOS/Linux, `cmd /c codragraph mcp` on Windows
#    (Node 22's spawn can't launch `.cmd` shims directly).
codragraph-codex

The codragraph-codex command is idempotent — re-run it after upgrades. A sidecar ~/.codex/.codragraph-managed.json tracks which entries the installer owns, so user-managed hooks/mcpServers are never overwritten.

What it does

Pre-tool hook: when Codex is about to Grep / Glob / Read / Bash, the hook calls codragraph augment <pattern> and prepends the graph context to Codex's next prompt. The hook reads bounded output from stdout or stderr because augment may write graph context on either stream.
Post-edit hook: after Edit / Write, performs a cheap metadata staleness check and a read-only working-tree dirty check. It never opens LadybugDB, never runs detect-changes, and never starts analyze in the background.
MCP server: registers a codragraph Codex MCP server (launching codragraph mcp, or cmd /c codragraph mcp on Windows) so Codex can call query / context / impact / feature_clusters / feature_context / cypher directly.

Configuration

API keys live in ~/.codragraph/config.json — the unified file the CLI, harness, and web app all use.

codragraph config set openai --api-key sk-...

License

Apache-2.0. You can use, modify, redistribute, bundle, and host this integration commercially, subject to the Apache-2.0 notice and attribution requirements.