← Projects·AI Engineer
eval-mcp-server
An MCP (Model Context Protocol) server that wraps the AI-output quality gate as a first-class tool any MCP-aware agent can call — exposes the 13-metric slop-evaluation gate over all three MCP primitives (Tools / Resources / Prompts) and proves the round-trip parity with a 20-case conformance suite.
Why this exists
MCP is the protocol that lets agents call tools from any model client (Claude Code, IDE plugins, custom orchestrators). A quality gate that lives only in one app doesn’t scale — an agent anywhere should be able to ask “is this draft ready to ship?” with the same confidence.
What it proves
- Conformance — every MCP primitive (Tools, Resources, Prompts) round-trips with the same input/output shape across 20 conformance test cases.
- Parity — calling the gate via MCP returns the same score as calling the in-process library directly.
- Composability — wraps the existing 13-metric literature-grounded slop evaluator (see
reflect-reviseandslop-scanner) as a drop-in for any MCP-aware client.
What’s transferable
The pattern — wrap an existing eval as an MCP server, prove round-trip parity, ship conformance tests — applies to any other evaluation you want to expose to a heterogeneous agent fleet.