Documentation

One endpoint. Thirteen tools.

Streamable HTTP MCP at https://faaast.ereneksi.com/mcp. OAuth 2.1 with PKCE, dynamic client registration, and API keys. Protocol versions: 2025-06-18, 2025-03-26, 2024-11-05.

Authentication

OAuth (recommended)

Point your client at the endpoint. It discovers authorization metadata from the 401 response and walks the full flow: dynamic registration → consent → authorization code + PKCE → rotating refresh tokens.

WWW-Authenticate: Bearer realm="faaast",
  resource_metadata="https://faaast.ereneksi.com/.well-known/oauth-protected-resource"

API key

Create a key in the dashboard and send it directly. Keys are prefixed faa_sk_, hashed at rest, and revocable instantly (cached up to 30s per edge location).

curl -s https://faaast.ereneksi.com/mcp \
  -H "Authorization: Bearer faa_sk_…" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Client setup

Claude Code

claude mcp add --transport http faaast \
  https://faaast.ereneksi.com/mcp

Cursor / VS Code / any JSON config

{
  "mcpServers": {
    "faaast": {
      "url": "https://faaast.ereneksi.com/mcp"
    }
  }
}

MCP Inspector

npx @modelcontextprotocol/inspector \
  --transport http \
  https://faaast.ereneksi.com/mcp

Raw JSON-RPC

{"jsonrpc":"2.0","id":1,"method":"initialize",
 "params":{"protocolVersion":"2025-06-18",
   "capabilities":{},
   "clientInfo":{"name":"curl","version":"1"}}}

Tools

Every tool returns content[0].text (compact, model-ready) plus structuredContent (JSON). Usage counters ride along in _meta.usage. All tools are read-only.

Tool	Arguments	What it does
`web_search` cost 1	`query` string required `count` integer = 6 `freshness` string: day\|week\|month\|year `sites` array `exclude` array `safe` boolean = true `deep` boolean = false `verbosity` string: terse\|normal = terse	Search the live web. Returns titles, canonical URLs, snippets and dates, fused across search backends and ranked for accuracy (not SEO). Use this first for factual questions, then fetch_url on the best source. Terse output by default.
`research` cost 3	`query` string required `sources` integer = 4 `max_chars_per_source` integer = 4000 `freshness` string: day\|week\|month\|year `deep` boolean = false	One call: search, read the best sources in parallel, and return only the passages that answer the question, with citations. This is the highest-accuracy tool for factual questions and comparisons. Cite the URLs it returns.
`fetch_url` cost 1	`url` string required `max_chars` integer = 12000 `include_links` boolean = false `allow_pdf` boolean = false `verbosity` string: terse\|normal = terse	Fetch one page and return clean, readable markdown (boilerplate, nav, scripts and ads removed). Respects robots.txt and AI opt-outs; never bypasses paywalls. Use after web_search to read the best source, or directly when you have a URL.
`batch_fetch` cost 1	`urls` array required `max_chars` integer = 6000 `include_links` boolean = false	Fetch up to 8 URLs in parallel and return clean markdown for each. Cheaper and faster than calling fetch_url repeatedly. Failures are reported per URL without aborting the batch.
`news` cost 1	`query` string required `count` integer = 8 `freshness` string: day\|week\|month = week `verbosity` string: terse\|normal = terse	Recent news and announcements, ranked by recency and relevance. Use for anything time-sensitive: releases, outages, funding, elections, prices.
`extract` cost 1	`url` string required `fields` object `json_ld` boolean = false	Pull exact fields out of a page with CSS selectors (title, price, author, links, table cells) or JSON-LD. Deterministic: no model, no guessing. Returns JSON.
`sitemap` cost 2	`url` string required `filter` string `limit` integer = 100	Discover a site's URLs from its robots.txt and XML sitemaps (handles sitemap indexes). Use to find the right page on docs sites, blogs and shops before fetching.
`wikipedia` cost 1	`query` string required `lang` string = en `full` boolean = false `max_chars` integer = 4000	Authoritative encyclopedia summaries from Wikipedia's API. Fastest reliable answer for definitions, people, places, history and science. Supports 300+ languages.
`github_search` cost 1	`query` string required `type` string: repos\|issues\|code\|users = repos `limit` integer = 6 `sort` string: stars\|updated\|relevance = stars	Search GitHub repositories, issues/PRs, code and users via the official API. Best for finding libraries, checking whether a bug is known, or locating real-world implementations.
`hackernews` cost 1	`query` string `sort` string: relevance\|date = relevance `limit` integer = 8	Search Hacker News stories and discussion via the Algolia API. Sharp signal on developer tools, launches and technical debates. No query returns the front page.
`arxiv` cost 1	`query` string required `limit` integer = 5 `sort` string: relevance\|submitted = relevance	Search arXiv preprints by title, abstract or author. Returns the paper ID, authors and abstract — the right source for ML, physics and math claims.
`package_info` cost 1	`name` string required `ecosystem` string: auto\|npm\|pypi = auto	Exact metadata for an npm or PyPI package: latest version, license, description, repo, dependency count, release date. Use instead of guessing versions.
`account_usage` cost 0		Current fair-use counters for this account: calls used today, per-minute headroom, reset time. Free to call. Use it to self-throttle before a large batch.

Examples

Search and read the best source

{"jsonrpc":"2.0","id":2,"method":"tools/call",
 "params":{
   "name":"web_search",
   "arguments":{"query":"cloudflare workers cpu limits",
                "count":5}}}

Grounded research with citations

{"jsonrpc":"2.0","id":3,"method":"tools/call",
 "params":{
   "name":"research",
   "arguments":{"query":"does deno support npm packages?",
                "sources":4}}}

Fair use & limits

Free2,000/day · 60/min · 12/s

Pro40,000/day · 400/min · 80/s

research3 units

batch_fetch1 unit per URL

account_usagefree

When a limit is hit the tool returns isError: true with the exact reset time. Nothing is silently dropped or slowed.

Why units, not tokens

A call costs the same whether the page is 2 KB or 40 KB. The meter lives in a Durable Object, so counting is atomic across the whole edge and adds ~1 ms — not a database round-trip.

What we will not do

No paywall bypass, no login walls, no CAPTCHA solving, no Google/Bing HTML scraping. Faaast obeys robots.txt, Content-Signal and X-Robots-Tag: noai, and identifies itself with a truthful user agent.

Errors

Tool-level failureresult.isError = true + plain-English text

Bad argumentsisError with the offending field

Unknown method/toolJSON-RPC -32601 / -32602

Missing or expired tokenHTTP 401 + WWW-Authenticate

robots.txt / AI opt-out403 with the exact rule that blocked it

Source, schema and deployment notes live in the repository. Questions: open an issue.

Ready?

Free account, 2,000 calls a day.

Create account