Fetch

Fetches web content from arbitrary URLs and automatically converts HTML/text responses into LLM-optimized formats (markdown, plain text, structured data). Uses HTTP client libraries with configurable headers and timeout handling to retrieve remote resources, then applies content extraction and normalization pipelines to strip boilerplate, extract main content, and format for efficient token consumption by language models.

Transforms raw HTML and text content into markdown format optimized for LLM consumption by removing unnecessary whitespace, normalizing heading hierarchies, converting HTML tables to markdown tables, and preserving semantic structure while minimizing token overhead. Uses HTML parsing libraries (likely html2text or similar) with custom post-processing rules to ensure output is both human-readable and token-efficient for language model analysis.

Registers the fetch and content-conversion capabilities as MCP tools that LLM clients can discover and invoke through the Model Context Protocol's JSON-RPC 2.0 interface. Implements the MCP server-side tool definition schema (including tool name, description, input schema with JSON Schema validation) and handles incoming tool call requests from clients, executing the appropriate fetch/conversion logic and returning results in the MCP response format with error handling for network failures, invalid URLs, and malformed requests.

Validates incoming URLs before fetching to prevent SSRF attacks, DNS rebinding, and access to sensitive internal services. Implements URL parsing to check for valid schemes (http/https only), validates against a blocklist of private IP ranges (127.0.0.1, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, localhost, etc.), and optionally enforces domain whitelisting. Rejects requests to file://, data://, and other non-HTTP schemes to prevent local file access and data exfiltration attacks.

Provides configurable HTTP client behavior through parameters for request timeouts, custom headers, user-agent strings, and connection pooling. Implements sensible defaults (e.g., 30-second timeout, standard user-agent) while allowing clients to override these settings per-request. Handles connection pooling and session reuse to improve performance for multiple sequential requests, and implements proper cleanup of resources to prevent connection leaks.

Implements comprehensive error handling for network failures (connection timeouts, DNS resolution failures, connection refused), HTTP errors (4xx, 5xx status codes), and content parsing errors. Returns structured error responses through the MCP protocol with error codes and human-readable messages, allowing clients to distinguish between transient failures (retry-able) and permanent failures (invalid URL, access denied). Implements exponential backoff retry logic for transient errors and provides detailed error context for debugging.