MrScraper MCP

MrScraper MCP Server for Web Scraping and Web Search

MrScraper MCP is the official MrScraper server for the Model Context Protocol. It connects AI assistants to page content and web search through two tools, plus an account check. Retrieve HTML, Markdown, or screenshots with scraperAPI, discover pages with webSearch, and check your plan and token usage with status.

Use the hosted server with OAuth sign-in, or run the open-source server locally with the @mrscraper/mcp npm package. See the MrScraper MCP documentation for setup instructions and client configurations.

The server exposes three tools:

scraperAPI  webSearch  status

Deployment options

The hosted server signs you in through your browser with OAuth 2.1, so there is no key to copy. Self-hosted and stdio setups use a MrScraper API key from app.mrscraper.com/api-tokens, which the hosted server also still accepts.

Connect to the hosted server

Point your MCP client at https://mcp.mrscraper.com/mcp with no credential. The server answers with an OAuth challenge, your client opens a browser, and you approve the connection on app.mrscraper.com.

claude mcp add --transport http --scope user \  mrscraper https://mcp.mrscraper.com/mcp

For clients configured by file, omit the headers block entirely:

{  "mcpServers": {    "mrscraper": {      "type": "http",      "url": "https://mcp.mrscraper.com/mcp"    }  }}

Use exactly this URL, including the /mcp path. It is the resource identifier access tokens are issued for, and a token minted for a different spelling of the address will be refused.

What you are approving

A tool called without its scope is refused with 403 insufficient_scope, and your client will offer to grant the missing permission.

scraperAPI submits a scraping job to the backend with POST /, which requires scrape:write. Connections authorized before this scope was advertised need to authorize again and grant it; scrape:read alone does not grant write access.

Disconnect it later under Connected applications in your MrScraper account. Disconnecting revokes the refresh tokens immediately; an access token already issued stops working within the hour.

Connect with an API key instead

An API key still works as a bearer token, which is what self-hosted deployments and non-OAuth clients use:

{  "mcpServers": {    "mrscraper": {      "type": "http",      "url": "https://mcp.mrscraper.com/mcp",      "headers": {        "Authorization": "Bearer YOUR_MRSCRAPER_API_KEY"      }    }  }}

Use your client's environment-variable or secret-storage feature so the key stays out of project files. An API key carries your full account authority, so no scope is enforced against it.

Codex

export MRSCRAPER_API_KEY="YOUR_MRSCRAPER_API_KEY"codex mcp add mrscraper \  --url https://mcp.mrscraper.com/mcp \  --bearer-token-env-var MRSCRAPER_API_KEY

Start Codex from an environment containing MRSCRAPER_API_KEY. Codex stores the environment-variable name in its MCP configuration.

Claude Code

claude mcp add \  --transport http \  --scope user \  --header "Authorization: Bearer YOUR_MRSCRAPER_API_KEY" \  mrscraper https://mcp.mrscraper.com/mcp

Claude Code also supports project and local scopes. Choose the scope that matches how broadly the connection should be available.

Copyable agent setup prompt

Connect MrScraper MCP to this agent. Detect the current MCP client and configure a Streamable HTTP server named mrscraper at https://mcp.mrscraper.com/mcp with no credential. The server uses OAuth 2.1, so the client will open a browser to sign in. If the client cannot do OAuth, fall back to a MrScraper API key from https://app.mrscraper.com/api-tokens stored through the client's environment-variable or secret-storage mechanism and sent as Authorization: Bearer <key>. Reload the MCP client when required, then list the tools and confirm that scraperAPI, webSearch, and status are the only available tools.

Choosing a tool

A discovery workflow is webSearch → scraperAPI. Both web tools advertise readOnlyHint: true, openWorldHint: true, and destructiveHint: false; status is read-only and not open-world. Authentication comes from the MCP connection; no tool takes an API token argument.

Migration from the previous tool set

Only scraperAPI, webSearch, and status are registered. The other old tool names are removed.

super_mode becomes super. home_page now maps to warmStart, and token_cap=0 is accepted. Restart the server and reconnect the MCP client after updating so it refreshes tools/list.

Changes in 0.3.0

webSearch now calls MrScraper's web search action on the Scraper API host and returns a ranked list of results. It gained num_results, and region must be a two-letter country code. The 0.2 options raw, format, language, page, and render_js are no longer supported: calls that still send them succeed, and the options are ignored. MRSCRAPER_SYNC_BASE_URL is gone; MRSCRAPER_FETCH_BASE_URL now covers both web tools.

Tool reference

The options and defaults of scraperAPI follow the Python SDK's scraper_api method. MCP tool names use camel case; their argument names use snake case.

scraperAPI

POST https://api.mrscraper.com/

Start with a known URL:

{  "url": "https://www.scrapethissite.com/pages/simple/"}

Request HTML, Markdown, and a screenshot together:

{  "url": "https://example.com",  "html": true,  "markdown": true,  "screenshot": true,  "super": true,  "home_page": true,  "geo_code": "ID",  "timeout": 120}

Omit browser_rendering or use null to let the service try without a browser and fall back to browser loading. true forces browser loading and false disables it. A screenshot always forces browser loading, even when browser_rendering=false. super selects real-device routing independently.

The JSON body contains url, maxRetries, and optional tokenCap. Page options are query parameters. For API keys, the request also includes the SDK-compatible token query parameter and Authorization/x-api-token headers. OAuth JWTs are sent only through Authorization.

webSearch

POST https://api.mrscraper.com

The JSON body carries action: "websearch", the query, and the optional country and result count. Ask for 20 results from the United Kingdom:

{  "url": "running shoes",  "region": "gb",  "num_results": 20}

A search results URL also works; its q parameter becomes the query.

status

GET https://api.app.mrscraper.com/api/v1/subscription-accounts

status takes no input. It returns the connected account's identity (name, email, and verification state), subscription status, token limit, token usage, remaining tokens, usage percentage, rate limit, auto-renewal, and renewal date. It copies only those fields, so the API key and Stripe identifiers in the raw account record never reach the client.

{  "kind": "mrscraper-cli-status-summary",  "source_endpoints": ["/subscription-accounts"],  "status_code": 200,  "data": {    "account": {      "subscription_status": "active",      "enterprise": false,      "token_usage": 61240,      "token_limit": 100000,      "token_remaining": 38760,      "usage_percent": 61.24,      "rate_limit": 0,      "rate_ttl": 0,      "auto_renew": true,      "ends_at": "2026-12-01T00:00:00.000Z",      "user": {        "name": "Ada",        "email": "[email protected]",        "verified": true      }    }  }}

An OAuth connection needs the account:read scope for status.

Response envelope

scraperAPI and webSearch return status_code, data, and headers in structuredContent and as JSON text. JSON bodies are parsed; other response bodies remain text. status returns the summary shown above.

A scraperAPI response carries the requested representations in data; screenshot is a base64-encoded JPEG and x-status-code is the target page's status:

{  "status_code": 200,  "data": {    "html": "<!DOCTYPE html>...",    "markdown": null,    "screenshot": null,    "token_usage": 3,    "runtime": 18.02,    "retry_count": 0,    "error": null  },  "headers": {    "content-type": "application/json",    "x-status-code": "200",    "token_usage": "3"  }}

A webSearch response lists the results in data.data, each with rank, title, url, description, and source:

{  "status_code": 200,  "data": {    "success": true,    "message": "Search completed",    "data": [      {        "rank": 1,        "title": "Web Scraping API | MrScraper",        "url": "https://mrscraper.com/",        "description": "Extract structured data from any website with AI.",        "source": "MrScraper"      }    ],    "tokenUsage": 1  },  "headers": { "content-type": "application/json; charset=utf-8" }}

Failures include error and set MCP isError=true. Network failures have status_code=null. Credential-bearing response headers and JSON metadata are filtered; network errors redact signed query credentials.

Manual testing: one tool at a time

scripts/test-tools.ts starts the compiled local MCP server over stdio, verifies that exactly the three tools are exposed, calls one tool, prints its response, and exits. No separate server terminal is needed. The runner loads .env, or reads MRSCRAPER_API_KEY / MRSCRAPER_API_TOKEN from your shell. Each tool call uses the live MrScraper service.

Run these commands individually from the repository directory:

1. Install dependencies:

   sh

   npm ci

2. Build the current server:

   sh

   npm run build

3. Set your API key:

   sh

   export MRSCRAPER_API_KEY="YOUR_MRSCRAPER_API_KEY"

4. Test scraperAPI:

   sh

   npm run test:manual -- scraperAPI --url "https://www.scrapethissite.com/pages/simple/"

5. Test webSearch:

   sh

   npm run test:manual -- webSearch --query "MrScraper web scraping"

6. Test status:

   sh

   npm run test:manual -- status

The webSearch sample asks for US results (region us). Each command prints Tools: scraperAPI, webSearch, status and PASS <tool> on success. Exit codes are 0 for success, 1 for a tool or connection failure, and 2 for invalid command arguments.

Pass a complete JSON object to test any supported options:

npm run test:manual -- scraperAPI --args '{"url":"https://example.com","markdown":true,"screenshot":true,"super":true}'

npm run test:manual -- webSearch --args '{"url":"web scraping","region":"id","num_results":20}'

To test an already running HTTP MCP server, append --target http://127.0.0.1:8000/mcp. Rebuild after source changes when testing local stdio. npm run test:manual -- --help lists runner options.

Run the server locally

Node.js 20 or newer is required.

git clone https://github.com/mrscraper-com/mrscraper-mcp.gitcd mrscraper-mcpnpm cinpm run build

Streamable HTTP

TRANSPORT=http npm start

The default endpoint is http://127.0.0.1:8000/mcp.

{  "mcpServers": {    "mrscraper": {      "type": "http",      "url": "http://127.0.0.1:8000/mcp",      "headers": {        "Authorization": "Bearer YOUR_MRSCRAPER_API_KEY"      }    }  }}

stdio

{  "mcpServers": {    "mrscraper": {      "command": "npx",      "args": ["-y", "@mrscraper/mcp@latest"],      "env": {        "MRSCRAPER_API_KEY": "YOUR_MRSCRAPER_API_KEY"      }    }  }}

Stdio credential precedence is MRSCRAPER_API_KEY, then MRSCRAPER_API_TOKEN.

Authentication and security

The HTTP transport is an OAuth 2.1 resource server. It publishes RFC 9728 protected resource metadata at /.well-known/oauth-protected-resource (and at the /mcp path-suffixed variant), and an unauthenticated request is answered with 401 and a WWW-Authenticate header naming that document, which is how a client discovers where to send you to sign in.

Access tokens are verified locally against the authorization server's published JWKS. A token must name this server in its aud claim, so a token issued for another MrScraper surface cannot be replayed here.

API keys are validated against the MrScraper account endpoint before tool execution. Either way, each HTTP request uses its own caller's credential; server environment credentials are never used on the HTTP transport.

The server filters credential-bearing response headers. Parsed JSON credential metadata and credentials embedded in generated curl commands are redacted. Page content remains available in data.

Browser-origin requests are accepted from trusted local origins, from claude.ai and chatgpt.com, and from exact origins configured through MRSCRAPER_ALLOWED_ORIGINS. Service-to-service MCP clients typically connect without an Origin header. The discovery documents are served ahead of origin checks, since any client must be able to read them before it has credentials.

Interactive widgets

webSearch carries a self-contained search results widget that lists the results with their titles, links, and descriptions. status carries an account widget with the plan, token usage, and a usage meter. Hosts without MCP Apps support display the normal JSON tool response.

Each widget is a self-contained document with its script and styles inlined, because strict host CSP blocks anything external. Each one declares an empty network allowlist, since it only renders data the tool already returned. Sources live in ui/; npm run build:widgets bundles them into src/widgets/bundles.generated.ts, which npm run build does for you.

Docker

docker build -f docker/Dockerfile -t mrscraper-mcp .docker run --rm -p 8000:8000 mrscraper-mcp

The image binds to 0.0.0.0. Apply the network controls appropriate for the deployment and keep bearer authentication enabled.

Environment variables

Troubleshooting

• Connection error: Confirm the server URL ends in /mcp and reload the MCP client after configuration changes.
• 401 Unauthorized: For OAuth, reconnect the server so your client runs the sign-in flow again. For an API key, confirm the bearer key for HTTP or the environment key for stdio.
• 403 insufficient_scope: The token lacks a scope the tool needs. Reconnect and approve the additional permission. For scraperAPI, the token must include scrape:write; for webSearch, it must include scrape:read.
• Signing in succeeds but tools fail: Confirm the server URL is exactly https://mcp.mrscraper.com/mcp; the token audience is bound to it.
• 403 Forbidden Origin: Add the exact browser origin to MRSCRAPER_ALLOWED_ORIGINS.
• Tool input error: Compare the call with the tool's parameter table and its defaults.
• Upstream timeout: Increase scraperAPI.timeout (up to 300 seconds) or webSearch.timeout when the target requires a longer request window.
• New tools are missing: Start a new client session and inspect tools/list for exactly scraperAPI and webSearch.

Development

npm cinpm run format:checknpm run lintnpm testnpm run buildnpm pack --dry-run

Smoke-test a running HTTP server:

npm run test:mcp -- \  --target http://127.0.0.1:8000/mcp \  --token "$MRSCRAPER_API_KEY"

npm run test:e2e runs both new tools sequentially against the live MrScraper service using the manual runner. Use -- --url URL --query QUERY to customize the targets, or -- --target http://127.0.0.1:8000/mcp to use HTTP.

MCP Registry

The hosted server's MCP Registry manifest is defined in server.json. It describes the Streamable HTTP endpoint and links to this repository and the setup documentation.

.github/workflows/publish-mcp-registry.yml publishes server.json as io.github.mrscraper-com/mrscraper-mcp. Run it from the Actions tab (Publish to MCP Registry → Run workflow) or push a v* tag that matches the version in server.json. It authenticates with GitHub OIDC, which grants this repository the io.github.mrscraper-com/* namespace; an interactive mcp-publisher login github works only for organization owners. The workflow fails if server.json and package.json have different versions.

See the official registry publishing guide for details. Registry versions are immutable, so use a new version in server.json when publishing changed metadata. Publishing this manifest registers the hosted server; it does not publish a new npm package.

License

MIT. See LICENSE.