quickstart: Claude Code

One command sets up wigolo and wires it into Claude Code:

npx wigolo init --agents=claude-code

What init does

Unattended by default — no prompts, safe in CI and dotfile scripts:

1. Warms the engine. Verifies the browser engine, the embeddings model, and the ML reranker (components download on first use; init pre-caches and verifies them), then prints the per-component setup report you see in the GIF.
2. Wires Claude Code. Registers the MCP server (user scope), adds a global-instructions block, installs the /wigolo command, and installs the wigolo skill packs — so Claude Code knows when and how to reach for search, fetch, crawl, research, and the rest.
3. Diagnoses. Runs the doctor's cold checks so a broken install fails loudly here, not mid-task.

Restart Claude Code afterwards and ask it to look something up — it now has ten local-first web tools with a persistent knowledge cache.

Variants

npx wigolo init --agents=claude-code,cursor,zed   # wire several agents at once
npx wigolo init                                   # engine only; point any MCP client at wigolo yourself
npx wigolo init --interactive                     # plain-text prompt flow
npx wigolo init --wizard                          # rich guided setup TUI
npx wigolo init --json                            # machine-readable summary

Supported agent ids: claude-code, cursor, zed, windsurf, codex, gemini-cli, opencode, vscode, and antigravity. (Cline gets the skill packs via wigolo skills add — wire its MCP entry by hand.) Undo everything with npx wigolo uninstall.

About the GIF

The recording runs the exact same init --agents=claude-code flow against this repo's local build (plus --skip-verify, which skips the post-install live smoke check) in a throwaway HOME, with model caches pre-warmed so you are not watching a progress bar. First-ever run on a fresh machine downloads the browser engine and models once — a few minutes on a normal connection — and every run after that looks like the GIF.