Safe Fetch

An injection-aware content fetcher, exposed over the Model Context Protocol.

Agents that read the open web read content nobody on your team wrote. A page or a post can carry a fake closing tag followed by "ignore previous instructions and ...". This server fetches a URL, strips it to text, runs defense-in-depth sanitization, caps the size, and caches the result. It is small on purpose. It is the reference implementation of a set of production hardening field notes, not a framework.

Why it exists

Most MCP servers are written for a demo: one caller, one happy path, no untrusted input, no cost ceiling. In production the failures cluster in three places, and none of them show up on day one:

• Untrusted content. A single tag stripper feels safe and is not. Unicode variants and unclosed tags walk straight through one regex.
• Cost. An unbounded response body or an oversized context is money spent on noise.
• Concurrency. Two tools touching one SQLite file throw database is locked.

This server answers all three in code you can read in five minutes.

What it does

fetch_clean(url, max_chars=2500) returns sanitized, size-capped text plus an audit of what was done. The defense is order, not cleverness:

1. strip the injection wrapper by its literal name, first
2. normalize unicode (NFKC) so homoglyph tags cannot hide
3. drop script and style bodies, then the generic tag strip
4. a second net for common instruction-override phrases

It caps input size with a [truncated] marker (if the model has to ignore most of the input, you are paying for nothing), clamps the per-call cap at the entry to a hard ceiling, and caches results in SQLite opened with WAL and busy_timeout so overlapping callers wait instead of crashing.

Install

pip install -e .

Run

As a standalone MCP server (stdio):

mcp-safe-fetch

Register it with an MCP client (for example, Claude Code) by pointing the client at the mcp-safe-fetch command. The server exposes one tool, fetch_clean.

Test

pip install -e ".[dev]"
pytest

The field notes behind it

The reasoning, with the production incidents that motivated each defense, is written up here: a short essay on MCP hardening (concurrency, prompt injection, cost) and what breaks after day 30. The sanitizer in src/mcp_safe_fetch/sanitize.py is the exact function from that write-up.

License

MIT.