A verified tool factory for AI agents. From any API surface to safe, verified tools any agent runtime can use — Go CLI + MCP server + agent skills + SQLite/FTS5 cache + dry-run-by-default policy + hash-verifiable proof artifacts.
30 seconds DEMO
$ gutenberg quick https://catfact.ninja/fact gutenberg quick: probing https://catfact.ninja/fact gutenberg quick: verdict = openapi-published (high) gutenberg quick: downloading spec from https://catfact.ninja/docs?api-docs.json gutenberg quick: generating tool at library/tools/catfact gutenberg quick: verifying catfact ← go build · cli-smoke · mcp-handshake · go test gutenberg quick: verify ok gutenberg quick: installing catfact → ~/.local/bin/catfact $ catfact fact ← Grade A · 100/100 · ~30 s end-to-end GET https://catfact.ninja/fact status: 200 OK { "fact": "In just 7 years, one un-spayed female cat and one un-neutered male cat...", "length": 121 }
One command. Any public API. A working tool surface for AI agents: CLI, MCP, skills, cache, policy and proofs. Verified by build, cli-smoke, MCP handshake, and Go tests. Nothing else does this.
What "quick" does, in 6 numbered steps
- probe the URL (
try) — classifies it asopenapi-published / json-endpoint / html-content / spa / lazy-rendered / anti-bot-challenge / http-error - fetch the spec (or seed a HAR if no spec is exposed; auto-infers baseUrl/schemas from real response bodies)
- generate Go CLI + MCP server + Claude skill + OpenClaw skill
- verify with build + cli smoke + MCP handshake + Go tests →
proofs/verification.json - install into
~/.local/bin/<slug> - expose heroes — zero-friction aliases auto-detected from the spec (e.g.
catfact fact,github meta,tvmaze shows)
Or pass a natural language intent and route to the catalog
$ gutenberg quick "top hacker news stories" gutenberg quick: searching catalog for "top hacker news stories" gutenberg quick: top match: hacker-news topstories-json (score=5) [...executes against the cataloged tool...]
Why Gutenberg
Most AI-agent infrastructure today gives you a model and tells you "bring your own tools". Gutenberg makes the tools — and makes sure they actually work before any agent calls them.
| Primitive | What it does |
|---|---|
forge / generate |
OpenAPI / GraphQL / HAR / OpenAPI-from-Postman/Insomnia/curl → Go CLI + MCP server + Claude/OpenClaw skill |
verify |
Runs go build, CLI smoke test, MCP handshake, go test — writes proofs/verification.json. No proof = no Grade A. |
try <url> |
Autopilot: classifies the site (openapi-published / json-endpoint / html-content / spa / lazy-rendered / anti-bot-challenge / http-error) and tells you the exact next command |
extract / scrape |
LLM-powered structured extraction (JSON-Schema validated, cache TTL, Anthropic/OpenAI/Ollama provider) + Firecrawl-style markdown scraping |
snapshot/replay |
Generated tools support deterministic record + offline replay for tests and CI |
vault |
Shared, AES-GCM-encrypted token vault across all generated tools |
lockfile + diff + upgrade + watch |
Pin spec hashes, diff semantically, regenerate while preserving // gutenberg:keep blocks, poll upstream for drift |
compare |
Call N tools on the same intent, see results side by side |
aggregator kind |
One recipe → a fan-out Go CLI over N JSON sources with merge.go + rank.go |
heroes |
Zero-friction aliases auto-detected (or curated via x-gutenberg-hero) — e.g. espn nba, github meta, tvmaze shows |
Installation
Gutenberg is currently installed from the GitHub repository. The npm package is not published yet, so do not use npx gutenberg for now.
Prerequisites for the full workflow:
- Node.js 20+
- Go installed on
PATH(go versionmust work), orGUTENBERG_GOROOT=/path/to/go - Git and curl
- Optional:
OPENAI_API_KEY,ANTHROPIC_API_KEY, or Ollama forextract/ LLM-routedrun
Install from source:
git clone https://github.com/JustVugg/gutenberg.git
cd gutenberg
npm install
node bin/gutenberg.js doctorUse it directly from the repo:
node bin/gutenberg.js search "top hacker news stories"
node bin/gutenberg.js try https://api.tvmaze.com/shows/1Or install the gutenberg command globally from this local checkout:
npm link gutenberg doctor
Generated tools are installed into ~/.local/bin. Make sure it is on your PATH:
export PATH="$HOME/.local/bin:$PATH"
For browser recording flows only, install a Playwright browser:
npx playwright install chromium
First Tool
The fastest end-to-end test is quick: it probes a URL, generates a tool, verifies it, and installs the generated CLI.
gutenberg quick https://catfact.ninja/fact catfact fact
If you did not run npm link, use node bin/gutenberg.js instead of gutenberg.
Generate from a public OpenAPI spec:
curl -fsSL -o /tmp/github.json \
https://raw.githubusercontent.com/github/rest-api-description/main/descriptions/api.github.com/api.github.com.json
gutenberg generate /tmp/github.json --out library/tools/github --name github --force --strict
gutenberg verify library/tools/github
gutenberg install library/tools/github
github meta
github call search/repos --param q=claude-code --select '[*].full_name'Probe a site you've never seen:
gutenberg try https://www.thetrainline.com/it # → "spa, use record" gutenberg try https://it.wikipedia.org/wiki/Bologna # → "html-content, use scrape/extract" gutenberg try https://api.tvmaze.com/shows/1 # → "json-endpoint, seed-har + import + generate"
What Gutenberg works on (honest table)
| Site shape | Recommended path | Success rate | Examples we've shipped |
|---|---|---|---|
| API with public OpenAPI | generate directly |
✅ ~95% | GitHub (1183 ops), Stripe (587), Sentry (209) |
| API without spec but JSON endpoints | seed-har → import-har → generate |
✅ ~90% | ESPN, TVMaze, Wikipedia REST, public JSON APIs |
| Plain HTML content site | scrape (+ --structured) or extract with schema |
✅ ~80% | Wikipedia and content/documentation pages |
| SPA with reachable XHR | record (Playwright) → import-har |
needs interactive triggering | |
| SPA + Cloudflare Turnstile / strong anti-bot | record --backend browserbase (paid) — or use an official partner API |
❌ ~5% with free tooling | thetrainline, lastminute, OpenTable, Subito.it |
Non-goals (what Gutenberg deliberately does NOT do)
- Bypass anti-bot protections. Cloudflare Turnstile, JS challenges, fingerprint-based bot detection: Gutenberg does not try to defeat them. When a site invests in stopping bots, that's a product decision we respect. For legitimate browser-driven scraping use the official Browserbase backend (
record --backend browserbase). - Solve CAPTCHAs. No 2Captcha-style integrations.
- Scrape against ToS. The README of every generated tool tells users to respect rate limits, robots.txt, and terms of service.
- Be a model. Gutenberg routes LLM calls (Anthropic / OpenAI / Ollama) but does not train or fine-tune.
Commands
gutenberg doctor Probe runtime, LLM providers, Browserbase gutenberg try <url> Autopilot — classify a site and pick the next command gutenberg discover <url> Look for a published OpenAPI/Swagger gutenberg seed-har <url> [<url>...] Fetch URL(s) via plain HTTP, emit HAR (no browser) gutenberg record <url> Playwright HAR recording [--storage-state state.json] [--backend browserbase --key $KEY --project-id $PROJ] gutenberg login <url> --out state.json Headed login flow, save storage-state for later record/extract gutenberg import-har <har> HAR → OpenAPI (with schema inference + auth sniffer) gutenberg import-graphql <source> GraphQL SDL / introspection → OpenAPI gutenberg generate <spec> --out <dir> OpenAPI → Go CLI + MCP + Claude skill (+ openclaw, optional) [--strict] [--min-score 0.7] [--default-header 'k: v'] [--policy policy.json] gutenberg generate <recipe> --kind aggregator --out <dir> gutenberg verify <project-dir> Runs build + cli-smoke + mcp-handshake + tests; writes proofs/ gutenberg install <project-dir> go build + symlink into ~/.local/bin gutenberg install starter-pack Install Hacker News, ESPN, Wikipedia, TVMaze, Open-Meteo, GitHub gutenberg scorecard <project-dir> Multi-dim scorecard with verification badges gutenberg publish <tool-dir>|--all Validate proofs + skills + OpenClaw; sync registry gutenberg registry sync Rebuild library/registry.json from real manifests gutenberg upgrade <project-dir> [--spec ...] Regenerate, preserve `// gutenberg:keep` blocks, auto go mod tidy gutenberg watch <project-dir> <spec-url> Poll spec URL, regenerate on drift gutenberg diff <old-spec> <new-spec> Semantic spec diff gutenberg extract <url> -p PROMPT -s SCHEMA LLM extraction with JSON-Schema validation + --cache 1h gutenberg scrape <url> [--structured] Main-content markdown (table detector for repeated patterns) gutenberg search <intent> Catalog discovery: top matching operations gutenberg run <intent> [--no-llm] LLM-routed (or top-match) execute against catalog gutenberg compare <tool-a> <tool-b> --op X Side-by-side call across tools gutenberg site build Rebuild web/data.js with scorecard + heroes
Generated tools share a uniform sub-command surface:
<tool> operations [--json] List operations <tool> heroes [--json] List auto-detected hero aliases <tool> <hero-alias> Shortcut to call --digest <tool> call <op-id> [options] Call with full control --param k=v --path k=v --header 'k: v' --data '{...}' --yes --json | --compact | --digest | --select 'data.events[*].name' --stream For SSE/NDJSON endpoints <tool> walk <op-id> --max N Auto-pagination (offset/limit, cursor, page) <tool> sync Cache read operations locally (SQLite/FTS5) <tool> search <query> Full-text over cached responses <tool> resources [query] Search projected resource rows in SQLite <tool> auth <status|client-credentials|device|pkce-start|pkce-finish|refresh> <tool> mcp Run as MCP stdio server
Catalog included
21 verified catalog entries with proofs/verification.json:
| Tool | Ops | Source | Verified checks |
|---|---|---|---|
| github | 1183 | GitHub OpenAPI ufficiale | build • cli • mcp • tests |
| stripe | 587 | Stripe OpenAPI ufficiale | build • cli • mcp • tests |
| sentry | 209 | Sentry OpenAPI ufficiale | build • cli • mcp • tests |
| espn | 4 | HAR-seeded | build • cli • mcp • tests |
| wikipedia-it | 5 | HAR-seeded (User-Agent default) | build • cli • mcp • tests |
| tvmaze | 5 | HAR-seeded | build • cli • mcp • tests |
| multi-demo | 2 sources | aggregator (ESPN + Open-Meteo) | build • cli • tests |
| dev-pulse | 2 sources | aggregator (GitHub + ESPN) | build • cli • tests |
Plus CatFact, CoinGecko, Dog API, Exchange Rates, Hacker News, npm registry, Open-Meteo, PokeAPI, public holidays, PyPI, RandomUser, REST Countries, and TVMaze quick. The complete source of truth is library/registry.json, generated by gutenberg registry sync.
Plus 16+ recipe templates in library/recipes/ for common sources (Hacker News, CoinGecko, PokeAPI, Open-Meteo, REST Countries, Nominatim, npm registry, PyPI, exchange rates, public holidays, Reddit, Spotify, Slack, …).
Output targets
--targets go,mcp,skill,openclaw controls what gets emitted next to the Go module:
go+mcp— Go binary + stdio MCP server (default)skill—skills/<slug>/SKILL.md(Claude Code skill format)openclaw—openclaw/<slug>/skill.{json,md}(agent-framework-neutral schema)
Architecture diff vs other open-source factories
Compared to site2cli (PyPI) and printingpress.dev (~82 CLIs, npm), Gutenberg uniquely ships:
- Verification proofs gate (no Grade A without
proofs/verification.json) gutenberg try <url>autopilot- Aggregator-as-recipe (declarative)
- Snapshot/replay deterministic mode (built into every generated tool)
- Shared AES-GCM token vault
lockfile+diff+upgrade+watchcross-tool compare- LLM-provider-agnostic extract (Anthropic / OpenAI / Ollama) + cache TTL
- HAR → OpenAPI with response schema inference and auth scheme sniffer
- Built-in JSONPath
--selectand--streamfor SSE/NDJSON
What we don't have yet: their catalog scale (82 vs 21) and packaged npm distribution. The catalog grows with recipes and gutenberg publish --all --fix-assets.
Run tests
node --test tests/*.test.jsLocal-network-dependent tests skip cleanly in sandboxes that deny 127.0.0.1 listeners. Browser tests remain opt-in (GUTENBERG_BROWSER_TESTS=1 to exercise live Playwright).
Environment variables
| Var | What |
|---|---|
ANTHROPIC_API_KEY |
Use Anthropic for extract / run |
OPENAI_API_KEY |
Use OpenAI for extract / run |
OLLAMA_HOST (default http://localhost:11434) + GUTENBERG_LLM_PROVIDER=ollama |
Use local or remote Ollama for extract |
BROWSERBASE_API_KEY + BROWSERBASE_PROJECT_ID |
Required for record --backend browserbase |
GUTENBERG_VAULT_KEY (32-byte hex) |
Encrypt the shared OAuth vault with AES-GCM |
GUTENBERG_TELEMETRY=1 + GUTENBERG_TELEMETRY_FILE |
Opt-in local JSONL log of operation calls (never sent anywhere) |
<TOOL_ENV>_SNAPSHOT_MODE=record|replay + <TOOL_ENV>_SNAPSHOT_DIR |
Deterministic HTTP record/replay per tool |
<TOOL_ENV>_AUTO_REFRESH=0 |
Disable automatic OAuth refresh in a generated tool |
License
MIT.
Status
v0.1.0 — public. Honest about its boundaries:
- ✅ Verified end-to-end on GitHub, Stripe, Sentry, ESPN, Wikipedia IT, TVMaze and the public-API starter catalog
- ✅ 21 catalog entries shipped, 16+ recipe templates
- ❌ Cannot bypass Cloudflare Turnstile / strong anti-bot. Use
record --backend browserbaseor an official partner API for that class of site.