AssertMCP
mid-run output assertions for AI agents — catch bad outputs before they post, commit, or email
npx -y github:Isco0819/assertmcpEver run into this?
Like promptfoo (5k+ GitHub stars, ProductHunt #1 for LLM eval tooling, paying teams) but promptfoo runs after the agent finishes. By the time CI flags a bad output your agent already sent the email, pushed the commit, or hit the API. There is no in-run way to assert output quality without calling an external LLM judge.
The fix
MCP server with three assertion tools agents call inline before any destructive action — schema check, safety check, content check — all local, zero external API, sub-10ms.
What it does
- assert_schema(json, schema) — AJV JSON Schema v7 validation gates agent output before it touches a database or downstream API
- assert_safe(text, blocklist) — regex and term matching blocks PII, brand violations, and forbidden content before agent posts or sends
- assert_contains(text, required[]) — confirms all required phrases are present before agent commits, deploys, or publishes