fastCRW MCP

Install

Works in

• Claude Code
• Cursor
• Codex
• Gemini CLI
• Windsurf
• OpenCode
• SourceCraft
• Koda
• Coddy

Manual install

npx -y crw-mcp@latest install

What it does

fastCRW is a single Rust binary with an engine for working with web pages. Give it a URL and it returns Markdown, HTML, a list of links, images or JSON matching your schema. Plain pages are fetched with an ordinary request, and for pages built by scripts the engine switches to a built-in lightweight browser or Chrome. The command `npx -y crw-mcp@latest install` puts both the `crw` MCP server and a set of skills into the agents it finds on your machine; without a key the server runs locally in embedded mode. The MCP server gives the agent `crw_scrape`, `crw_crawl` with progress polling through `crw_check_crawl_status`, `crw_map`, `crw_search`, `crw_extract` and `crw_parse_file`. The map tool collects addresses from the sitemap and links without downloading content. A crawl runs breadth first from a start URL and follows `robots.txt` by default. `crw_parse_file` turns a base64 PDF into Markdown. By default a response is cut to 15,000 characters so it does not flood the agent's context, and `maxLength: 0` lifts the cut. The repository holds thirteen skills built around one ladder of steps: search first, then scrape a single page, map, crawl, parse a PDF, extract by schema and watch for changes. Separate skills cover moving from Firecrawl by swapping the base URL and standing up your own server in Docker. The same operations are available from the `crw` command line, from Python and Node.js SDKs and over REST. The cloud needs a key from fastcrw.com, while local mode is free.

Who it is for. For developers and analysts whose agents in Claude Code, Cursor or Codex read sites, documentation and search results, and for anyone who wants the page collector on their own server.

Example request

Collect the documentation from docs.example.com as Markdown: first show how many pages there are, then fetch the first fifty

MCP

Transport

stdio

Authentication

not required

Limitations

• The engine and MCP server are AGPL-3.0, while the Python and TypeScript SDKs are MIT. For embedding in a closed product the author asks you to write by email.
• Cloud mode is paid by credits: an account gets 1000 free credits once, not monthly. The key can be stored in `~/.config/crw/config.toml` and is not copied into each agent's settings.
• `crw_extract` and the `json` and `summary` formats of `crw_parse_file` need a connected language model: your own provider key or a model on the server.
• To keep agent settings untouched, set `CRW_NO_AGENTS=1`; without it the installer with a key registers the server in Claude Code, Cursor, Codex, Gemini CLI, OpenCode and Windsurf by itself.
• Available features depend on the deployment: the `/v1/capabilities` address returns the list. Crawl and map follow `robots.txt`, and respecting site rules is up to the user.

How to disable. Remove the `crw` entry from your agent's MCP configuration and the `crw*` skill folders from its skills directory. The `crw` binary placed by the install script sits in `/usr/local/bin` or the folder from `CRW_INSTALL_DIR`; deleting it is enough. The key and settings are kept in `~/.config/crw/config.toml`.

README in short

The repository holds the engine sources split into Rust crates (server, CLI, MCP, search, rendering, page comparison), a Python SDK, thirteen skills, plugin manifests for Claude Code, Codex and Cursor, speed benchmarks with their methodology, and blog articles. The front page opens with a one-command install and a table of operations, followed by a choice between the cloud and your own server and links to the docs.