MCPJam features
Everything MCPJam does to help you trust your MCP server: inspector, OAuth debugging, JSON-RPC logging, evals, and more.
MCPJam Inspector
Open-source playground for MCP servers: connect a server, run its tools, compare models side by side, read the full waterfall trace. Free and npx-installable.
MCP OAuth & EMA debugger
The MCP OAuth Debugger visualizes every step of the auth flow so you see exactly where it breaks, with spec conformance checks and EMA Cross-App Access tracing.
Cross-client testing
Generate an eval suite from your MCP server's tools, run it with a real model, and get accuracy, latency, and diagnosed failure patterns, not only protocol checks.
CI/CD actions
Gate every PR on how models use your MCP server. Wire MCPJam evals into GitHub Actions, track accuracy over time, and catch behavior regressions early.
Test Agent Skills + MCP locally
Load Anthropic Agent Skills alongside your MCP servers and watch a real model use both. Skills viewer plus an LLM playground, running locally with one npx command.
Chatboxes
Hosted, shareable UAT environments that mirror the major AI hosts. Testers break things safely, and their sessions become replayable regression tests.
Synthetic acceptance testing with Swarm
Turn loose AI personas that act like your users to acceptance-test your MCP server at scale, with multi-turn journeys across every client, before launch.
Frequently asked questions
The Inspector. Connect your MCP server, call its tools, and watch a real model use them — most problems surface in that first session. Evals, CI/CD gates, and cross-client testing are what you add once you want that same check to run without you.
The core is open source and free: the client, Inspector, core CLI and SDK, local evals, and conformance checks, runnable locally or in your own CI with one npx command and no signup. Paid plans add shared credits, higher eval limits, SSO, and support — pricing carries the current split.
Yes. MCPJam is SDK, framework, and deployment agnostic: if an MCP client can connect to your server, MCPJam can test it, using the same config shape as an mcp.json file. It is also strictly pre-production — it exercises your server the way a client would and never instruments live traffic.
Yes. The CLI and SDK run the same checks headlessly, and CI/CD eval gates wire them into GitHub Actions so a pull request fails on a behavior regression rather than a user finding it. The hands-on features and the automated ones are the same engine.