Firecrawl

Extracts and converts web page content from a single URL into either Markdown or HTML format through the firecrawl_scrape tool. The MCP server accepts a URL and optional parameters (format, headers, wait time), forwards the request to Firecrawl's backend via the @mendable/firecrawl-js client library, and returns structured content with metadata. The tool handles transport-agnostic communication through stdio, SSE, or cloud transports depending on deployment configuration.

Extracts content from multiple URLs in a single request through the firecrawl_batch_scrape tool, which submits an array of URLs to Firecrawl's batch processing pipeline. The server forwards the batch to the backend, which processes URLs in parallel (respecting rate limits), and returns an array of content objects with per-URL status and metadata. This capability leverages Firecrawl's internal job queue and credit pooling to optimize throughput for multi-page research tasks.

Abstracts communication with both cloud-hosted and self-hosted Firecrawl instances through a unified @mendable/firecrawl-js client interface. The server accepts a FIRECRAWL_API_URL environment variable to specify a custom endpoint (for self-hosted deployments) or uses the default cloud endpoint. All 8 tools transparently work with either deployment model, allowing operators to switch between cloud and self-hosted without code changes. This pattern enables cost optimization (self-hosted for high volume) and data sovereignty (self-hosted for sensitive data).

Discovers and extracts URLs from a base domain using the firecrawl_map tool, which crawls the target site's structure and returns a list of discovered URLs. The tool uses Firecrawl's crawler to traverse links, respect robots.txt, and build a URL graph; it returns a flat array of URLs found on the domain, useful for understanding site structure before targeted scraping. The MCP server forwards the base URL and optional depth/limit parameters to Firecrawl's mapping engine.

Submits a crawl job for a domain and polls its status asynchronously through firecrawl_crawl and firecrawl_check_crawl_status tools. The firecrawl_crawl tool initiates a background crawl job (returning a job ID), and firecrawl_check_crawl_status polls the job's progress, returning status (running/completed/failed), progress percentage, and partial results. This pattern enables long-running crawls without blocking the MCP client, leveraging Firecrawl's job queue and background processing.

Extracts structured data from web content using LLM-powered extraction through the firecrawl_extract tool. The tool accepts a URL and a JSON schema or natural language description of desired fields, submits the request to Firecrawl's backend (which fetches the page and uses an LLM to extract matching fields), and returns structured JSON matching the provided schema. This capability combines web scraping with semantic understanding, enabling extraction of complex nested data without regex or CSS selectors.

Performs web search and automatically scrapes top results through the firecrawl_search tool. The tool accepts a search query, submits it to a search backend (Google, Bing, or Firecrawl's internal index), retrieves top results, and optionally scrapes content from matching URLs. The MCP server returns an array of search results with URLs and optionally extracted content, enabling agents to research topics without pre-existing URL lists.

Implements automatic retry logic with exponential backoff for transient failures across all Firecrawl operations. The MCP server wraps tool calls with a retry mechanism configured via environment variables (FIRECRAWL_RETRY_MAX_ATTEMPTS, FIRECRAWL_RETRY_INITIAL_DELAY, FIRECRAWL_RETRY_BACKOFF_FACTOR, FIRECRAWL_RETRY_MAX_DELAY). On failure, the server waits for an exponentially increasing duration before retrying, capping the delay at a maximum. This pattern handles rate limiting, temporary network issues, and backend unavailability transparently.

Tracks Firecrawl credit consumption across all operations and emits alerts when usage crosses configurable thresholds. The MCP server queries Firecrawl's account API to fetch remaining credits, compares against warning and critical thresholds (FIRECRAWL_CREDIT_WARNING_THRESHOLD, FIRECRAWL_CREDIT_CRITICAL_THRESHOLD), and logs alerts or raises exceptions when thresholds are breached. This capability enables proactive cost management and prevents unexpected quota exhaustion.

Supports multiple communication transports for MCP clients through configurable protocol selection. The server can operate in stdio mode (for CLI/desktop clients), SSE local mode (for local web integration), or SSE cloud mode (for multi-tenant SaaS). Transport selection is determined by environment variables (SSE_LOCAL, SSE_CLOUD) and the @modelcontextprotocol/sdk's transport abstraction. This enables the same server code to serve different deployment architectures without modification.

Validates incoming MCP tool calls against strict JSON schemas and marshals arguments into typed objects before forwarding to Firecrawl. Each of the 8 tools (scrape, batch_scrape, map, crawl, check_crawl_status, extract, search) defines a schema specifying required/optional arguments, types, and constraints. The MCP server validates incoming calls against these schemas, rejects invalid requests with detailed error messages, and passes validated arguments to the Firecrawl client. This pattern ensures type safety and prevents malformed requests from reaching the backend.