Crawlbrulee

🍮 crawlbrulee mcp

[npm] [license]

EU-native web scraping for AI agents & developers.

plug crawlbrulee into your agent. the official mcp server for crawlbrulee gives mcp-aware agents — Claude Code, Codex, Cursor, Claude Desktop — native tools to scrape pages, map sites, run background jobs, and check usage. one call turns any url into clean markdown, screenshots, metadata and links.

• everything runs in the EU. the fetch, the render, the cache and your result never leave EU servers. the proxy exit is the one hop you choose: pick an EU exit and nothing leaves at all. gdpr-aligned, with a data processing agreement.
• output made for models. markdown with ads, popups and cookie banners removed and the links kept, ready for the prompt. full-page screenshots can come back as tiles sized for an image model.
• the hard parts, handled. headless Chrome when a page needs it, rotating proxies with country selection, automatic retries, ad and cookie-banner removal, caching, background jobs and signed webhooks.
• start free. 750 credits, no credit card.

get a free api key → dashboard.crawlbrulee.com

the server:

• npx-runnable — zero install.
• wraps the @crawlbrulee/sdk under the hood; this mcp is just a thin protocol adapter.
• stdio transport for terminal-based agents.
• strict, fully-described tool schemas — agents see what every parameter does without reading docs.

this readme covers the mcp server itself — its tools and how to wire it into a host. for how the api behaves — endpoints, parameters, and error semantics — please see our api docs.

install

1. get a key. if you don't have a crawlbrulee account, sign up. the free plan needs no credit card. then, in the dashboard menu, open tokens and click create token.

2. put it in your environment. add it to your shell profile (for example ~/.zshrc), then open a new terminal and start your agent from it. don't paste the key into your agent's chat: anything in the chat stays in its history.

export CRAWLBRULEE_API_KEY="cwbl_..."

3. add the server, with no key in its config. in Claude Code:

claude mcp add crawlbrulee -- npx -y @crawlbrulee/mcp

other stdio hosts run the same command: npx -y @crawlbrulee/mcp. a few of them (Codex, Cursor, Windsurf, Gemini CLI, Claude Desktop) need the key or the variable named in their config file, and VS Code and the Claude Code plugin can ask for the key in a masked prompt. the file and the entry for each one: mcp setup in the docs.

if a key stops working (invalid_credentials), it's expired, revoked or mistyped: create a new one the same way, replace the old one, and start your agent again.

configuration

the mcp reads the env var on first tool invocation — not at startup — so a typo in your config surfaces as a clear tool-error message rather than the server failing to come up. see authentication for how the api consumes keys.

tools

scrape

fetch a single url and return the requested content (markdown, cleaned html, raw html, links, images, screenshot, page metadata).

input — only url is required; everything else has sane defaults.

{  "url": "https://example.com",  "extract": {    "markdown": true,    "links": true,    "screenshot": { "type": "full_page", "device_mode": "desktop" },  },  "require_js": false,  "proxy": "basic",  "cleanup": { "ads_and_popups": true, "exclude_selectors": ["nav", "footer"] },  "cache": { "max_age": 3600 },  "location": { "locale": "en-US", "country": "US" },  "zero_data_retention": false,}

zero_data_retention (boolean, default false) keeps the result out of the shared cache; anything stored to deliver it is kept for 24 hours, then deleted. it adds 1 credit and must be enabled for your organization. scrape_async takes it too. see zero data retention.

named values with extract.elements. to pull a few values from a page — prices, titles, links — give each one a name and a CSS selector, instead of reading the whole markdown. they come back in elements under the same names. all: true returns every match as a list, and fields reads named values inside each match (:scope reads the match itself). it costs no extra credits. scrape_async takes it too.

{  "url": "https://shop.example.com/books/",  "extract": {    "elements": {      "heading": "h1",      "books": {        "selector": "article.book",        "all": true,        "fields": {          "title": { "selector": "h3 a", "output": "attribute", "attribute": "title" },          "price": ".price",        },      },    },  },}// → "elements": {//     "heading": "Books",//     "books": [{ "title": "The Night Garden", "price": "€18.90" }, ...]//   }

a name with no match is null (or [] with all: true). limits and selector rules: see targeted extraction.

your own headers and cookies with headers and cookies. send your own HTTP headers and cookies to the site, for example an Authorization header or a fixed User-Agent that a site has on its allowlist. headers go only to the origin of url (or its https version, for an http url) and cookies only to its host, never to another site on a redirect. up to 10 headers and 20 cookies. a request with headers or cookies never reads from or writes to the cache, and costs the same as a normal fresh scrape. scrape_async takes them too.

{  "url": "https://app.example.com/orders",  "headers": { "Authorization": "Bearer ...", "X-Tenant-Id": "acme" },  "cookies": [{ "name": "session_id", "value": "a1b2c3" }],}

[!WARNING] header and cookie values pass through your chat with the AI, so the AI host may keep them in its history. use short-lived tokens or a test account, and never a password you can't change.

your own User-Agent can lower the success rate, and the response then carries the custom_user_agent warning. never put secrets in the User-Agent. the limits and which headers can't be set: see headers and cookies.

output — full scrape result. page metadata (title, OG tags, etc.) is returned under metadata. extracted images are returned as absolute urls — query strings are preserved, and relative srcs are resolved against the page url. screenshots are returned as signed download urls the agent can fetch separately. in rare cases a screenshot can't be captured: when you requested other outputs too, the screenshot field is simply left out while the rest is still returned — but a screenshot-only call that can't deliver errors instead (unsupported_screenshot_output, HTTP 422, when the content type can't be screenshotted) and isn't billed. the result also carries page_status_code and a top-level response_meta.usage block:

{  "url": "https://example.com",  "requested_url": "https://example.com",  "page_status_code": 200, // the site's own HTTP status for the final page  "markdown": "...",  "metadata": { "title": "Example Domain" },  "response_meta": {    "usage": {      // total_credit_cost = engine_credit_cost × proxy_multiplier + screenshot_slicing_credit_cost + zero_data_retention_credit_cost      "total_credit_cost": 1,      "engine_credit_cost": 1, // http 1, browser 3, screenshot 5, cache 0      "proxy_multiplier": 1, // basic 1, advanced 5      "screenshot_slicing_credit_cost": 0, // 1 when the screenshot was split into slices, otherwise 0      "zero_data_retention_credit_cost": 0, // 1 when zero_data_retention added its credit, otherwise 0      "engine": "http", // "http" | "browser" | "screenshot" | "cache"      "proxy": "basic", // resolved tier actually used: "basic" | "advanced" (never "auto")    },  },}

a page the site served is a result, not an error. page_status_code is the HTTP status the site answered with for the final page, after redirects. a 404, 410 or 503 page comes back with its content and its status here, so check page_status_code before you trust the content: a 404 means the markdown is the site's "not found" page. 2xx and 4xx pages are billed, except 403, 407, 408, 429 and 451; 5xx pages are never billed. when the site can't be reached at all, the tool returns a target_unreachable error instead (see errors).

alongside response_meta.usage, the result surfaces any non-fatal warnings — stable string codes an agent can switch on. an outsized page is truncated rather than refused, and the code names which part was cut:

two more codes are not about size. screenshot_unavailable means a screenshot was asked for, but the page came back from the http engine without one; the rest of the result is still there. custom_user_agent means the request set its own User-Agent header; this can lower the success rate.

and if you requested an extract that doesn't apply to the content type (e.g. metadata of a JSON file), the field name comes back in an unsupported_fields list — with the rest of the payload still returned.

every input field, its default, and its constraints are documented under the scrape endpoint — with extraction, screenshots, proxies & location, and caching covering the individual blocks.

scrape_async

submit a scrape job to run asynchronously and get back a job_id immediately, instead of holding the connection open. use this for long-running scrapes (heavy js rendering, full-page screenshots of long pages); for a quick one-shot fetch prefer the synchronous scrape tool. then poll scrape_status until the job is done and fetch the page with scrape_result.

takes the same input as scrape plus an optional per-job completion webhook:

{  "url": "https://example.com",  "extract": { "markdown": true },  "webhook": {    // Endpoint that receives one signed `scrape.complete` POST when the job    // finishes. http/https (HTTPS required in production), max 2048 chars.    "url": "https://hooks.example.com/cwbl",    // Opaque correlation object echoed back verbatim in the delivery's    // `data.metadata`. Serializes to at most 2048 bytes.    "metadata": { "ref": "order-42" },  },}

output — { "job_id": "..." }.

when a webhook is attached, we deliver a single signed scrape.complete POST to your endpoint once the job reaches a terminal state, with your metadata echoed under data.metadata and the job's usage under data.response_meta.usage — so you can react to completion (and track cost) without polling. verify the X-Cwbl-Signature header with the sdk's verifyWebhookSignature (configure the signing secret in the dashboard under account → webhooks).

the job lifecycle is documented under async scrape; the delivery contract and payload shape under webhooks, with the signature scheme in webhook verification.

scrape_status

look up the current lifecycle status of an async job: pending, running, done, or failed (with an error message when failed). once the job is done the response also carries a response_meta.usage block (total_credit_cost and its parts, billed engine, resolved proxy tier). a cache hit is represented by engine: "cache". a job whose page the site served ends done even when that page is a 404 — read page_status_code in the result. poll until done, then call scrape_result.

{ "job_id": "..." }

scrape_result

fetch the extracted content of a completed async job — the same result shape as the synchronous scrape tool (including page_status_code, metadata, elements and response_meta.usage). errors if the job is still pending/running, so check scrape_status first.

{ "job_id": "..." }

map

build (or fetch a cached) link-map for a website. combines sitemap discovery with homepage link extraction. use this to enumerate a site before scraping selected pages. each link is just { url }.

max_urls (default 5000, max 100000) is a discovery budget, not a trim at the end — discovery stops as soon as that many urls are found, so a smaller value is a faster, cheaper crawl. limit (default 5000, max 10000) only pages the answer.

returned urls are normalized the same way scrape normalizes its returned url, so map-then-scrape stays on one host. results are ordered with the most useful links first.

the response's response_meta carries pagination, truncation, and a usage block (total_credit_cost = engine_credit_cost × proxy_multiplier + zero_data_retention_credit_cost, billed engine, resolved proxy tier). map has no screenshot slicing and no page_status_code, since it reads many files, not one page. a map that found nothing because the site answered only with statuses we don't bill (a 5xx, for example), or not at all, is empty and free.

{  "url": "https://example.com",  "sitemap_only": false,  "types": { "internal": true, "external": false, "internal_subdomains": true },  "max_urls": 5000,  "page": 1,  "limit": 1000,  "zero_data_retention": false,}

zero_data_retention works as it does on scrape (see zero data retention).

a map stopped by your own max_urls returns exactly that many links with response_capped: false — the signal that the site has more is truncation.discovery_cap_reason:

{  "truncation": {    "storage_capped": false,    "response_capped": false,    "total_before_max_urls": 5000,    "total_detected_before_storage_cap": 5000,    "discovery_capped": true, // discovery stopped before reading every sitemap file    "sitemaps_skipped": 3, // files skipped or only partly read    "discovery_cap_reason": "max_urls", // retry with a higher max_urls  },}

discovery_cap_reason is one of max_urls, time, file_budget, depth, file_size, unread_files, or null when nothing stopped discovery. only max_urls is a limit you can raise from the request. unread_files means a sitemap file the site publishes could not be read at all this time — often temporary, so asking again later can return more. time, file_budget, depth and file_size mean the site itself is big, slow or deep, and a retry will not help.

see the map endpoint for discovery rules and pagination semantics.

usage

returns the current billing-cycle snapshot: total / used / available credits, used quota percent, max concurrency, and cycle reset timestamp. takes no arguments. what a call costs, and how credits are counted, is documented under credits & pricing.

whoami

returns the organization name, token name, and truncated token preview for the configured api key. useful for confirming which account is in use before credit-consuming operations.

errors

every tool returns an mcp error result (isError: true) when the api call fails. the error text follows a stable format:

[<errorName>] <message> (HTTP <status>)

a few codes add a short next step after that, e.g. target_unreachable and zero_data_retention_not_enabled. branch on the errorName code, not on the rest of the text.

agents can branch on the errorName code. the set comes from the sdk's ApiErrorName union plus two synthetic codes added by this mcp (missing_api_key, internal_error):

the api docs carry the canonical error reference — every error name, what causes it, and how to recover.

development

pnpm installpnpm typecheck   # tsc --noEmitpnpm lint        # eslintpnpm test        # vitest runpnpm build       # tsup → dist/index.js with shebangpnpm verify      # all of the above

run the built mcp locally:

CRAWLBRULEE_API_KEY=cwbl_... node ./dist/index.js

it will block waiting for an mcp client on stdio. combine with the MCP Inspector for interactive debugging.

docs

this readme covers the mcp server itself — installing it, wiring it into a host, and the tools it exposes. for how the api behaves — endpoints, parameters, and error semantics — the api docs are canonical. the mcp guide covers host setup in more depth.

part of the crawlbrulee toolkit

one api, many ways to call it:

• js/ts sdk — @crawlbrulee/sdk (the sdk this mcp wraps)
• python sdk — crawlbrulee on pypi
• cli — npx crawlbrulee
• mcp server — @crawlbrulee/mcp (this one)
• agent skills — for skills-aware coding agents

docs: crawlbrulee.com/docs · dashboard: dashboard.crawlbrulee.com

license

Apache-2.0