MCP testing platform
The testing platform for MCP servers.
One open-source tool to inspect the protocol, run a real model against your tools, step through OAuth, and compare behavior across ChatGPT, Claude, and Cursor. It's the Postman for MCP, running locally with no subscription, plus a built-in tunnel when you want to test in a real client.
npx @mcpjam/inspector@latestWhat is the best MCP testing platform?
The best MCP testing platform covers more than protocol checks: it runs a real model against your server, scores how the model uses your tools, debugs the OAuth flow, and compares behavior across clients like ChatGPT, Claude, and Cursor. MCPJam does all of that in one open-source tool that runs locally. It's a Postman for MCP, with no subscription and built-in tunneling when you need a real client.
What makes a testing platform, not just an inspector
A protocol inspector answers one question: does the server respond correctly? That's necessary, but it's the smallest part of knowing an MCP server works. The moment a real model starts calling your tools, the questions that actually break production appear: which tool did it pick, why did the arguments come out wrong, where exactly did OAuth 401, and does any of it behave the same in ChatGPT as in Claude.
MCPJam is built around those questions. It covers protocol validation like a classic inspector, then adds the model-in-the-loop layers a platform needs: a Playground to run real models, an OAuth Debugger, cross-client testing, a JSON-RPC trace, and Evals that quantify behavior across many requests. It's open source and runs locally, the coverage of a platform without the lock-in of a cloud.
Everything the platform covers
Protocol inspection
Connect any local or remote server and confirm it initializes, lists tools, resources, and prompts, and returns valid schemas, the job a reference inspector does, built in.
Model-in-the-loop Playground
Run a real model against your server and watch which tool it selects, with what arguments, and how it recovers, the half no protocol check can see.
Evals
Generate a suite from your tools, run it with a real model, and get accuracy, latency, tokens, and diagnosed failure patterns across many realistic requests.
OAuth Debugger
Step through each request in the OAuth exchange between client, MCP server, and authorization server to find the exact point a 401 breaks the flow.
Cross-client testing
The same server behaves differently in ChatGPT, Claude, Gemini, and Cursor. Test across the clients it will actually meet in production, side by side.
CI/CD eval gates
Wire the CLI and SDK into GitHub Actions to run evals on every PR, so a behavior regression fails the gate instead of reaching users.
Start testing in one command
No install and no ChatGPT subscription. Point the inspector at any local or remote MCP server and start testing, with a built-in tunnel when you need to test in a real client.
# Any local or remote MCP server
npx @mcpjam/inspector@latestIs there a Postman for MCP?
"Postman for MCP" is the shorthand teams reach for, and the analogy holds where it matters: one place to connect a server, exercise its tools, inspect requests and responses, and share how it should behave. The difference is that an MCP server isn't a plain API; the caller is a language model, and correctness depends on how that model chooses and uses tools. So MCPJam adds the parts Postman was never built for: a real model in the loop, evals that score tool selection, and an OAuth debugger for the MCP auth flow.
If you're comparing tools by name, the honest breakdowns live on the comparison pages: MCPJam versus the official MCP Inspector, versus Postman's MCP support, and versus adjacent platforms like mcp-use and Alpic.
Open source, used in the open
Frequently asked questions
The right tool depends on how far you need to go. A protocol inspector is enough to confirm a server responds; MCPJam adds the model-in-the-loop layers (Playground, cross-client testing, an OAuth debugger, and evals) in one open-source tool that runs locally.
It plays the same role (one place to connect, exercise, and inspect an MCP server) but adds what an API client can't: a real model calling your tools, evals that score that behavior, and an OAuth-flow debugger built for MCP.
Yes. MCPJam's core is free and open source and runs locally with npx @mcpjam/inspector@latest and no subscription, plus a built-in tunnel for a real client when you need it. Team and Enterprise plans add scale.
This page is the platform overview; the how-to guide walks the full workflow step by step: protocol validation, behavior testing, OAuth, cross-client checks, and CI/CD gates.
Yes. The MCPJam CLI and SDK wire into GitHub Actions or any pipeline to run eval suites on every PR with pass/fail exit codes, so a regression blocks the merge.
Test your MCP server on the platform built for it.
Protocol validation, a real agent in the loop, OAuth debugging, and evals. Open source, running locally, in about two minutes.
Open source core · no credit card · ~2 min
Keep exploring
The full testing workflow, from local inspection to CI/CD.
The local inspector at the core of the platform.
Where the reference inspector stops and MCPJam continues.
Score how models use your server across realistic requests.
Read every JSON-RPC call and reproduce any failure.
Find the exact step where MCP auth breaks.