draco

draco scrape url [options]draco search query [options]draco serve [options]draco interact url [options]draco mcp

draco is a fast, native-Rust web scraper designed as a lighter, self-hostable alternative to Firecrawl and Browserbase. It fetches pages with a browser-faithful TLS/JA4 fingerprint, extracts main content as clean Markdown (and metadata), and can escalate client-rendered SPAs through an in-process V8 isolate — without shipping a headless Chrome fleet or Node runtime.For ordinary HTML, a scrape is typically a single fingerprinted fetch plus parse. Output includes Markdown (headings, absolutized links, lists, code fences, GFM tables; nav/footer/ads stripped), metadata (title, description, language, canonical, favicon, Open Graph / Twitter / article tags, status), and optional trace / timing steps. Default scrape exit codes: 0 success, 1 error, 2 unsupported, 3 needs_browser (e.g. a real JS challenge wall such as Cloudflare).Beyond one-shot scraping, Draco provides search (parallel multi-engine SERP fetch merged by reciprocal-rank consensus), serve (persistent HTTP daemon with a Firecrawl-compatible REST API on port 3002 by default, including scrape, map, crawl, batch scrape, discover, search, and interact), interact (stateful page sessions with cookie persistence and JS exec), and mcp (Model Context Protocol tools over stdio or the daemon's POST /mcp).JSON extraction escalates through static embedded state (__NEXT_DATA__, JSON-LD, Nuxt), Next.js build-id replay, then runtime fetch/XHR interception in a sandboxed V8 isolate with no host-capability bindings. Defaults respect robots.txt and rate limits; use is intended for public data and properties you are permitted to scrape.

--format fmt

Output format for scrape: markdown (default), html, raw-html, links, json, endpoints, or both (markdown+json). May be repeated where supported.

Emit the full result envelope as JSON (markdown + metadata + trace).

Pretty-print JSON output.

Select a path from extracted JSON (with --format json).

Route traffic through a proxy (e.g. socks5://127.0.0.1:9050).

Delay between requests for politeness.

Per-request timeout.

Cap escalation tier (0/1/2). 2 enables V8 runtime interception and SPA render-then-Markdown (default when built with tier2).

Disable main-content extraction heuristics.

Wait window related to runtime capture / hydration.

Cap the Tier-2 capture window duration.

Do not honor robots.txt (use only when you have permission).

Allow replaying state-changing requests during runtime modes that would otherwise stub them.

Print isolate diagnostics (exceptions, console errors, brokered fetches, capture close reasons).

For search: maximum number of results (default typically 5, range-capped).

For serve: bind address, port, and concurrent scrape budget.

For interact: run a one-shot JavaScript expression/body against the page.

scrape url

Fetch a URL and emit Markdown (default) or other formats. Primary CLI entrypoint.

Multi-engine web search over Draco's stealth HTTP stack; optional scrape of each hit.

Run the persistent Firecrawl-compatible HTTP API (default http://127.0.0.1:3002).

Open a one-shot or REPL session against a page: run JS, navigate, scrape live DOM.

Speak MCP over stdio for agent clients (Claude Desktop, editors, orchestrators).

Draco does not defeat JavaScript challenge walls (Cloudflare, DataDome, and similar). A genuine interstitial short-circuits to exit code 3 / needs_browser. A normal HTTP 200 behind a CDN is not treated as a challenge.Tier-2 isolation relies on an in-process V8 context with no host bindings (no filesystem or arbitrary network from page JS except engine-brokered fetches). It is not a full OS jail; older --no-jail / --strict-sandbox flags are accepted but inert. You remain responsible for target sites' terms of service and applicable law.Platforms with full support include Linux x86_64-gnu and macOS aarch64-darwin. Lean builds without default features omit V8 and the serve daemon.

draco is written in Rust as a single-binary, self-hostable Firecrawl alternative with native Markdown extraction and optional in-process V8 for SPA hydration and API discovery. It is dual-licensed MIT OR Apache-2.0. The OS process jail used in earlier releases was retired in v0.14 in favor of isolate-only containment for throughput.

curl(1), wget(1), crawlie(1), trafilatura(1), shot-scraper(1)

Source code