Optional Setup¶

These features are not required for the toolkit to work. Each adds a specific capability.

Codex CLI — automatic Phase 3 delegation¶

When Codex CLI is installed, /implement Phase 3 (implementation) delegates to Codex automatically. Claude falls back silently if Codex is unavailable.

npm install -g @openai/codex
codex   # authenticate once via browser

After installing, say /upgrade so the toolkit registers it.

Crawl4AI — clean web fetching¶

Routes all WebFetch calls through a local Crawl4AI container. Outputs clean Markdown, caches pages for 24 hours — the same URL is never fetched twice in a day. Requires Docker.

docker run -d -p 11235:11235 --name crawl4ai --shm-size=1g unclecode/crawl4ai:latest

Without this, the toolkit falls back to Claude's built-in WebFetch.

Local LLM — zero API cost for mechanical tasks¶

Routes mechanical tasks (boilerplate, docstrings, simple refactors, unit test scaffolding) to a local model via LM Studio. Results are cached for 24 hours.

Install LM Studio and download a model
Start the local server in LM Studio (default port: 1234)
Edit toolbox/tools/local-llm/config.json to point at your model

Claude checks LM Studio availability at session start. If unreachable, it falls back to a Haiku sub-agent silently.