readable-mcp

Turn any URL into clean, LLM-ready Markdown. A production MCP server: rate-limited, retried, cached, and SSRF-safe.

LLMs choke on raw HTML — nav bars, ad markup, and tracking scripts burn tokens and bury the actual content. Worse, a naïve "just fetch the URL" tool is a server-side request forgery (SSRF) hole waiting to be pointed at 169.254.169.254 or your internal network. readable-mcp solves both: it extracts the real content as Markdown/text/cleaned-HTML, and it does the fetching the way a production service would — validated, rate-limited, retried with backoff, cached, and observable. This is not a weekend demo; every outbound request assumes the network is hostile.

Features

Production qualities first — these are what make it safe to point at the open web:

• SSRF-safe fetching — resolves the hostname and refuses any address in private, loopback, link-local, or reserved space, so localhost, raw private IPs, and public hostnames that resolve inward are all blocked.

• Token-bucket rate limiting — a single shared limiter (default 5 req/s, burst 10) gates every outbound request, including inside batch calls.

• Exponential-backoff retries — retries timeouts, connection errors, 429, and 5xx with jittered backoff, and honors Retry-After on 429. Never retries other 4xx (those are caller errors).

• Explicit timeouts — separate connect/read and a hard total ceiling; a request can never hang forever.

• TTL caching — in-memory cache keyed by normalized URL + format (default 900s); cache hits are flagged from_cache=true.

• Structured typed errors — tools never raise to the client; failures come back as a typed {code, message, retryable}.

• First-class partial-success batch — one bad URL never fails the batch; each URL gets its own success/error slot.

• JSON observability logs — one structured line per call (request id, tool, host, latency, cache hit/miss, outcome) — never response bodies or secrets.

And the capability itself:

• Clean content extraction — trafilatura strips boilerplate and pulls title, author, and publish date; output as Markdown, plain text, or cleaned HTML.

Related MCP server: intercept-mcp

Architecture

 ┌────────────┐   extract_url / extract_batch / get_stats
 │ MCP client │──────────────────────────────────────────────┐
 │ (Claude,   │                                               │
 │  Cursor…)  │                                               ▼
 └────────────┘                                     ┌───────────────────┐
                                                     │  server.py (tools)│  thin orchestration
                                                     └─────────┬─────────┘
                                                               │
        ┌──────────────┬───────────────┬────────────┬─────────┴────────┐
        ▼              ▼               ▼            ▼                  ▼
 ┌─────────────┐ ┌────────────┐ ┌──────────┐ ┌──────────────┐ ┌──────────────┐
 │ validation  │ │ rate_limiter│ │  cache   │ │ http_client  │ │  extractor   │
 │  + SSRF     │→│ token bucket│→│ TTL (hit │→│ retry+timeout│→│ trafilatura  │
 │  guard      │ │  acquire()  │ │  / miss) │ │  httpx       │ │  → Markdown  │
 └─────────────┘ └────────────┘ └──────────┘ └──────────────┘ └──────┬───────┘
                                                                      ▼
                                                          ┌───────────────────────┐
                                                          │ typed ExtractionResult │
                                                          │   or ExtractionError   │
                                                          └───────────────────────┘

Validation runs before anything touches a socket; the rate limiter and cache sit in front of the network; every failure is funneled into a typed result.

Production handling, not a demo

Four patterns pulled straight from the source — the reason this is portfolio-grade.

1. SSRF guard that resolves before it trusts

A string check on the hostname is not enough: an attacker can register a public domain whose DNS points at 169.254.169.254. We resolve first, then check every resolved IP against the non-public ranges.

# validation.py
host = parts.hostname
try:
    literal = ipaddress.ip_address(host)      # raw-IP host?
    candidates = [str(literal)]
except ValueError:
    candidates = _resolve_addresses(host)     # otherwise resolve DNS

for ip in candidates:
    if not _is_public_ip(ip):                 # private/loopback/link-local/reserved
        raise ValidationError(
            ErrorCode.BLOCKED_HOST,
            f"Refusing to fetch {host!r}: resolves to non-public address {ip}.",
        )

Why it matters: this is the difference between a fetch tool and an internal-network proxy for whoever controls the input. It's the first thing a security-minded reviewer checks.

2. Async token-bucket rate limiter

One shared bucket throttles every outbound request — including the concurrent fetches inside extract_batch — so the server stays a polite citizen under load.

# rate_limiter.py
async def acquire(self, tokens: float = 1.0) -> None:
    while True:
        async with self._lock:
            now = asyncio.get_event_loop().time()
            self._refill(now)
            if self._tokens >= tokens:
                self._tokens -= tokens
                return
            wait = (tokens - self._tokens) / self._rate
        await asyncio.sleep(wait)   # sleep outside the lock

Why it matters: tokens refill continuously and the wait happens outside the lock, so many coroutines can share one limiter without deadlocking or over-drawing the bucket.

3. Retry with backoff that honors Retry-After

Transient failures (429, 5xx, timeouts, connection errors) are retried with jittered exponential backoff; a 429 with a Retry-After header is obeyed exactly. Other 4xx are returned immediately — retrying a 404 is just wasted requests.

# http_client.py
if status in _RETRYABLE_STATUS:
    if not is_last:
        retry_after = _parse_retry_after(response.headers.get("Retry-After"))
        delay = retry_after if retry_after is not None else _backoff_delay(
            attempt, self._settings.retry_base_delay
        )
        await self._sleep(delay)
        continue
    raise last_error
if status >= 400:
    raise FetchError(ErrorCode.HTTP_ERROR, ..., retryable=False)  # caller error, no retry

Why it matters: backoff with jitter prevents thundering-herd retries, and honoring Retry-After is what keeps you from getting hard-blocked by the upstream.

4. Structured errors — the client never sees a stack trace

Every failure path converges on one typed shape. Tools return it instead of raising, so a bad URL degrades gracefully instead of crashing the tool call.

# server.py
def _error(url: str, code: ErrorCode, message: str, *, retryable: bool) -> ExtractionError:
    return ExtractionError(
        url=url, error=ErrorDetail(code=code, message=message, retryable=retryable)
    )

Why it matters: the calling LLM can branch on error.code and error.retryable programmatically, and in a batch one bad URL simply occupies its own error slot.

Quickstart

With uv (recommended):

git clone https://github.com/tommypj/readable-mcp.git
cd readable-mcp
uv sync                 # install
uv run readable-mcp     # run the server over stdio

With pip:

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .
readable-mcp

The server speaks the MCP stdio transport, so it's normally launched by an MCP client (below) rather than run by hand.

Use it in Claude Desktop / Claude Code

Add this to your claude_desktop_config.json (mirrors examples/claude_desktop_config.json):

{
  "mcpServers": {
    "readable": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/readable-mcp", "run", "readable-mcp"]
    }
  }
}

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Claude Code: claude mcp add readable -- uv --directory /absolute/path/to/readable-mcp run readable-mcp

• Cursor: add the same mcpServers block under Settings → MCP.

Restart the client, then ask: "Extract the main content of https://example.com as markdown."

Tools reference

extract_url(url, output_format="markdown")

Fetch one URL and return its main content plus metadata.

• url — an http/https URL. Private/loopback/link-local hosts are refused.

• output_format — "markdown" (default), "text", or "html" (cleaned).

• Returns ExtractionResult or ExtractionError.

• Error codes: INVALID_URL, BLOCKED_HOST, FETCH_TIMEOUT, HTTP_ERROR, EXTRACTION_FAILED, RATE_LIMITED.

// example result (trimmed)
{
  "url": "https://example.com/",
  "final_url": "https://example.com/",
  "title": "Example Domain",
  "author": null,
  "published": null,
  "word_count": 17,
  "output_format": "markdown",
  "content": "This domain is for use in documentation examples...",
  "from_cache": false,
  "fetched_at": "2026-06-23T10:01:22.481+00:00"
}

extract_batch(urls, output_format="markdown")

Extract up to 10 URLs concurrently; partial success is first-class.

• urls — list of URLs (max 10 processed; the rest return TOO_MANY_URLS).

• Returns BatchResult with requested / succeeded / failed counts and a per-URL results list (each a result or a typed error).

{ "requested": 2, "succeeded": 1, "failed": 1,
  "results": [
    { "url": "https://example.com/", "title": "Example Domain", "word_count": 17, ... },
    { "url": "http://127.0.0.1/", "error": { "code": "BLOCKED_HOST", "message": "...", "retryable": false } }
  ] }

get_stats()

Return a lightweight operational snapshot — uptime, requests served, cache hit/miss + hit rate, and current config. No sensitive data.

{ "version": "0.1.0", "uptime_seconds": 124.3, "requests_served": 6,
  "cache_hits": 2, "cache_misses": 1, "cache_hit_rate": 0.6667,
  "rate_limit_rps": 5.0, "cache_ttl_seconds": 900 }

Configuration

All settings are environment variables prefixed READABLE_MCP_ (or a .env file — see .env.example). No secrets are required.

Testing

uv run pytest          # 49 tests, fully offline (network mocked with respx)
uv run ruff check .    # lint
uv run ruff format .   # format

Tests cover the production paths directly: SSRF/validation cases, token-bucket timing under concurrency, cache hit/miss/expiry, retry-on-429/503 + Retry-After + give-up + no-retry-on-404, extraction against saved HTML fixtures, and the tool happy/error/partial-success paths.

Design decisions

• trafilatura for extraction — best-in-class boilerplate removal with built-in title/author/date detection and native Markdown output; markdownify is the fallback when no main body is detected, so the caller still gets usable content rather than an error.

• Resolve-then-check SSRF — checking the literal host is insufficient; we resolve DNS and validate every returned IP so a public hostname can't tunnel to private space. Literal-IP hosts skip DNS and are checked directly.

• In-memory TTL cache (not Redis) — an MCP server is a single local process per client; an in-process TTLCache under an asyncio.Lock gives the hit-rate win with zero external dependencies. Swappable behind the cache.py boundary if a shared cache is ever needed.

• Errors as values, not exceptions — tools return typed ExtractionErrors so the model can branch on code/retryable and a batch can carry mixed success/failure. The server is designed to never crash on bad input.

• Shared rate limiter + concurrency cap for batches — the token bucket bounds throughput while a semaphore bounds simultaneous sockets, so a 10-URL batch is both polite and bounded.

• max_retries = total attempts — named for the common env-var convention, but semantically the attempt ceiling (default 3 ⇒ 2 retries). Documented here to avoid the off-by-one ambiguity.

• Unknown output_format — returned as a typed EXTRACTION_FAILED error with a clear message rather than silently coercing, so callers learn about the mistake.

License

MIT © Dan Tomescu. See LICENSE.