CASE STUDY — MODEL CONTEXT PROTOCOL
readable-mcp: safe web content for AI tools.
A production-oriented MCP server that turns permitted web pages into clean Markdown while treating network access, failure behavior, and operational limits as first-class engineering concerns.
Signals
What this project proves.
Coverage49 tests
Repeatable behavior across core paths.
SecuritySSRF-safe
Resolve, validate, redirect-check, and pin.
ProtocolMCP
Structured tool surface for AI clients.
LicenseMIT
Public implementation available for review.
Challenge
Letting an AI fetch URLs creates a real security boundary.
A useful extraction tool must handle hostile or malformed URLs, private network targets, redirects, oversized responses, timeouts, and upstream failures without exposing internal systems or returning ambiguous errors.
Approach
Design the network boundary before the happy path.
The server validates resolved addresses, rechecks redirects, pins approved destinations, limits requests, retries transient failures, caches safe results, and returns typed errors that an AI client can reason about.
Result
A reusable pattern for dependable MCP integrations.
The public implementation demonstrates the production safeguards that also apply when connecting an AI client to private APIs, databases, documents, or internal tools.
Stack
Tools and operating surface.
FastMCP
HTTP safety controls
SSRF protection
Typed errors
Rate limiting
Retries & caching
Automated tests
build pattern
Need this kind of system in your own stack?
Bring the workflow and the constraints. I’ll map the smallest production system worth shipping.
Book a scoping call