Perf

You can integrate perf-mcp into your MCP client to automatically fact-check, repair, and validate AI outputs before they reach users. This server provides multi-channel verification to catch hallucinations, fix malformed JSON, correct errors, and route requests to the most suitable model or workflow based on task complexity.

How to use

You integrate perf-mcp into your MCP client to gain automated quality assurance for LLM outputs. Use perf-mcp to verify factual accuracy, repair JSON structures, and correct semantic or instruction drift in real-time. The server offers specialized tools for detection, validation, correction, and routing to the best-performing models.

How to install

Prerequisites you need before installing perf-mcp:

• Node.js and npm installed on your development machine.

Follow these steps to configure perf-mcp in your MCP client. Use the exact MCP server configuration shown here to ensure proper startup and authentication.

How to configure in your MCP client

{
  "mcpServers": {
    "perf": {
      "command": "npx",
      "args": ["-y", "perf-mcp"],
      "env": {
        "PERF_API_KEY": "pk_live_xxx"
      }
    }
  }
}

Environment variables

The following environment variables are used by perf-mcp. Provide them in your client configuration to enable secure access and proper routing.

Pricing and licensing

Pricing is based on the number of tool calls. The free tier includes 200 verifications. Subscriptions and per-verification pricing apply for higher usage. This server is released under the MIT license.

Tools and capabilities

perf-mcp offers a set of tools to improve AI output quality. You can use these tools to detect hallucinations, validate JSON against a schema, perform general corrections, and route requests to the most appropriate model.

Security considerations

Protect your API key and only expose necessary endpoints to your MCP client. Use environment-specific keys for testing and production keys for live environments.

Notes

If you need to test configurations, start with the free tier and verify that perf-mcp properly validates and corrects outputs before enabling higher traffic. Ensure your client clears sensitive data after each verification cycle.

Troubleshooting

If perf-mcp fails to start, verify that Node.js and npm are installed, and confirm that your MCP client can access the configured command and environment variables. Ensure the API key is correctly set in the environment. Check for network restrictions that could block API requests.

License

This software is released under the MIT license.

Tools overview

perf_verify: Detects and repairs hallucinations using multi-channel verification including web search, natural language inference, and cross-reference checks. Returns a corrected output with confidence estimates.

perf_validate: Validates LLM outputs against a given JSON schema and auto-repairs violations such as wrong types, missing fields, and hallucinated properties.

perf_correct: General-purpose output correction that identifies the error type (hallucination, schema violation, semantic mismatch, or instruction drift) and applies the appropriate fix.

perf_chat: Routes requests to the most suitable model among multiple options based on task complexity, using an OpenAI-compatible format.

Available tools

perf_verify

Detects and repairs hallucinations using multi-channel verification (web search, NLI models, cross-reference) rather than relying on a single model.

perf_validate

Validates a JSON string against a target schema and auto-repairs violations such as wrong types or missing fields.

perf_correct

Classifies the error type and applies the appropriate correction, including hallucinations, schema violations, or semantic drift.

perf_chat

Routes requests to the optimal model automatically based on task complexity, selecting from multiple models.