The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Fetch MCP Server listing page.
One click adds this to your local Yaw MCP config so it's available in every Yaw Terminal session. Or install manually below.
A comprehensive HTTP fetch MCP server for AI assistants. Bring-your-own client: runs as a stdio MCP server so any MCP-compatible client (Claude Code, Claude Desktop, Cursor, mcph, …) can fetch web content safely.
| Tool | What it does |
|---|---|
http_get / http_head / http_options | Bare HTTP requests with headers, auth, timeout, size cap, retry |
http_post / http_put / http_patch / http_delete | Write-method HTTP with JSON or raw body |
fetch_html_to_markdown | GET a page and convert to clean markdown (3–8× smaller than raw HTML) |
fetch_html_to_text | GET a page and convert to plain text with block structure preserved |
fetch_reader | Reader-mode extraction — isolates the article body and returns title + markdown |
fetch_meta | Extract <head> metadata: title, description, OpenGraph, Twitter cards, JSON-LD, feeds, icons |
fetch_links | Extract every outbound link, resolved to absolute URLs, classified internal/external |
fetch_sitemap | Parse sitemap.xml (including gzipped and sitemap-index chaining) |
fetch_feed | Parse an RSS 2.0 or Atom 1.0 feed into entries |
fetch_robots | Parse a site's robots.txt, return the verdict for a given path & user-agent |
SSRF protection is on by default. The server refuses requests to:
127.0.0.0/8, ::1)10/8, 172.16/12, 192.168/16)169.254/16, fe80::/10) — including the cloud metadata endpoint 169.254.169.254100.64/10)fc00::/7)::ffff:0:0/96) re-checked against the IPv4 rules, and likewise the IPv4 embedded in IPv4-compatible (::/96), IPv4-translated (::ffff:0:0:0/96) and 6to4 (2002::/16) addresses2001::/32), NAT64 (64:ff9b::/96, 64:ff9b:1::/48), site-local (fec0::/10) and discard-only (100::/64) IPv62000::/3), and the non-global blocks inside it (3fff::/20, 2001:2::/48, 2001:10::/28)168.63.129.16http/https schemes (file://, gopher://, javascript:, …)localhost and any *.localhostEvery redirect hop is re-validated against all of the above -- scheme, literal IP and localhost* -- before it is dialed, so a 302 from a public host to http://127.0.0.1, http://169.254.169.254 or ftp://… is caught. For hostnames, DNS is resolved once per hop, every returned address is checked, and the verified IP is pinned into the HTTP dispatcher so the subsequent TCP connection dials that exact address — closing the DNS-rebinding TOCTOU window. Authorization, Cookie and Proxy-Authorization headers are stripped on cross-origin redirects.
The model chooses tool arguments, and the model is not trusted: a prompt injected through a fetched page can ask for anything. So the per-call allow_private_hosts: true opt-in is refused unless the operator enables it by launching the server with FETCH_MCP_ALLOW_PRIVATE_HOSTS=1 (true/yes/on also work; any unrecognised value is treated as off and named on stderr). With the variable set, a call still has to ask: requests without allow_private_hosts stay guarded.
Only set it where the model may legitimately reach your internal network -- a local dev box, not a cloud VM with a metadata endpoint.
Requires Node 22.19 or newer when running on Node (the floor of its HTTP client, undici 8; the launcher refuses older Nodes with a clear message instead of crashing). Under oam, which the launcher prefers when installed, the Node version does not matter.
Add to your client's MCP config (usually claude_desktop_config.json or ~/.claude.json):
Or via mcph:
http_get, http_post, http_put, http_patch, http_delete, http_head, http_optionsCommon parameters:
| Field | Type | Default | Meaning |
|---|---|---|---|
url | string | — | Absolute URL |
headers | object | — | Custom request headers |
timeout_ms | int | 10000 | Timeout for each attempt, covering DNS, every redirect hop and the body. Retries get a fresh budget; a whole call is capped at 5 minutes. Cancelling the tool call stops the request |
max_bytes | int | 5242880 (5 MiB) | Truncate body if larger |
max_redirects | int | 5 | Redirect hops allowed |
retries | int | 0 | Retry count on 408/425/429/5xx with backoff (honors Retry-After) |
user_agent | string | @yawlabs/fetch-mcp/<v> | User-Agent override |
basic_auth | {username,password} | — | Injects Authorization: Basic … |
bearer_token | string | — | Injects Authorization: Bearer … |
allow_private_hosts | bool | false | Opt this call into loopback / private / link-local targets. Refused unless the operator set FETCH_MCP_ALLOW_PRIVATE_HOSTS=1 (details) |
decode_text | bool | auto | When unset, auto-detects by response Content-Type (text for text/*, JSON, XML, JS, form-urlencoded; binary otherwise). Set explicitly true to force text decoding, false to force base64 in body_base64. |
Body-capable tools (POST/PUT/PATCH/DELETE) also take:
| Field | Type | Meaning |
|---|---|---|
body | string | Raw request body |
body_json | any | Structured body — encoded as JSON, Content-Type: application/json set automatically |
content_type | string | Overrides Content-Type |
Response shape:
fetch_html_to_markdownGET the URL, strip scripts/styles/iframes/svg/canvas plus <nav>, <footer>, <aside>, convert to atx-headed markdown with fenced code blocks and dash bullets. Intended for feeding pages into an LLM without blowing the context budget.
fetch_html_to_textSame fetch, but emits plain text with block-level structure preserved as newlines. Useful when the model doesn't need markdown formatting.
fetch_readerIsolates the main article body using, in order: <article>, <main>, itemprop="articleBody", common CMS class names (post-content, entry-content, etc.), then <body> as fallback. Returns:
fetch_metaGET a URL and return its head metadata without downloading the full body (caps at 2 MiB by default):
fetch_linksGET a page and return every <a href> with text, resolved to absolute URLs. Respects <base href>. Skips #, javascript:, mailto:, tel:, data:, file:. Each link is classified internal or external vs. the page host. Optional filter/dedupe/limit.
fetch_sitemapFetch a sitemap.xml or sitemap-index and return the URL list:
Gzipped .xml.gz sitemaps are auto-decompressed. max_depth controls how many levels of sitemap-index to follow (default 1). Setting max_depth: 0 on a sitemap-index returns the index's childSitemaps list without fetching any child (useful to discover structure cheaply). Partial failures — one child sitemap 500s while others succeed — are returned under warnings rather than aborting the whole call. max_sitemaps (default 50, max 1000) caps how many sitemap documents one call fetches, index included; children past it are listed under childSitemaps with a warning. The whole call is capped at 5 minutes. max_bytes (default 20 MiB) applies to each sitemap, and to the decompressed size of a gzipped one; a sitemap larger than that is an error (a warning for a child) rather than a silently partial URL list.
fetch_feedParse an RSS 2.0 or Atom 1.0 feed:
fetch_robotsFetches <origin>/robots.txt, parses it, and returns:
A /robots.txt that returns 404 means there are no rules, so the verdict is allowed: true. The response keeps the same shape: status: 404, matchedRule: null, crawlDelay: null, sitemaps: [], rawRobotsText: "", plus the note. Any other non-2xx status is returned as an error.
The parser follows Google's rules: longest match wins, * is a wildcard segment, $ anchors the end of the path, specific user-agent group beats the * wildcard group when the UA matches (comparison uses the length of the actually-matched agent token, not the group's first agent). Allow beats Disallow on equal-length ties.
Tests spin up a local loopback HTTP server on 127.0.0.1:0 to exercise the real request/response path — no mocking of HTTP. SSRF tests verify that the default-deny still applies to that local server unless the operator gate is on and the request opts into allow_private_hosts.
MIT © Yaw Labs