Fetch

Fetch MCP is the simplest, smallest MCP in the catalog: one tool, `fetch`, that takes a URL and returns the HTTP response body. It handles redirects, sets a reasonable User-Agent, and returns content with the appropriate content-type. Many forks add convenience parameters like `max_length` (truncate the response), `raw` (skip HTML-to-markdown conversion), and `start_index` (pagination). The official MCP from the Anthropic / community team is the most-installed across coding-agent contexts.

We use Fetch MCP everywhere as the "second half" of a search workflow. Brave Search returns 10 URLs; Fetch MCP reads the top 3 in parallel. We also use it for API exploration — "what does the GitHub API return for `/user/repos`" — without writing any client code. Third use case: pulling RSS / Atom feeds for content monitoring. Token cost is whatever the page is; a typical blog post in markdown is ~3-8k tokens. Most forks of Fetch MCP convert HTML to markdown automatically, which both makes the content readable for the model and dramatically reduces tokens vs raw HTML.

Compared to writing a curl command in Bash MCP, Fetch wins on structured behaviour (timeouts, redirect handling, content-type detection) and on safety (no shell injection). Compared to Playwright/Puppeteer for scraping, Fetch wins on speed (sub-second vs multi-second) but loses on any page requiring JavaScript execution. The honest trade-off: many modern sites (Twitter, LinkedIn, anything with a heavy SPA) return near-empty HTML on a plain fetch — the content lives in JS-rendered DOM. For those, fall back to Playwright/Puppeteer.