Summary
A second opinion that is not just another model's opinion.
Verification here goes through web search and NLI models rather than one grader prompt, and the JSON repair covers the failure that actually breaks pipelines: output that parses cleanly but does not fit the schema.
What it is
An MCP server that checks and repairs LLM output before it reaches a user. Verification runs multi-channel — web search, NLI models, cross-reference — rather than asking a second model to grade the first. It also validates JSON against a schema and routes chat requests across more than 20 models.
What you get
- Text checked for hallucinations, fabricated facts and unsupported claims, with the correction and a confidence figure
- Generated JSON validated against a schema and auto-repaired: malformed enums, wrong types, missing fields, hallucinated properties
- The error type classified — hallucination, schema violation, semantic inconsistency, instruction drift — and the matching fix applied
- Requests routed to the model that suits the task, across GPT-4o, Claude, Gemini, Llama and 20+ others, in OpenAI-compatible format
Requirements
An API key in PERF_API_KEY, from the Perf dashboard; it arrives with a batch of free verifications and needs no card. It runs from npm with npx.
Setup effort
One command plus a key — npx -y perf-mcp, then supply credentials