Inspecting a server by hand tells you it works in the thing you are holding. Running the same test cases across sixteen client configurations and a range of models tells you where it breaks for someone else, which is the failure mode that reaches users. Evals with expected tool calls turn that into something a pipeline can gate on, and the OAuth debugger covers the other half — the part that fails before any tool is ever called.
A testing and evaluation platform for people building MCP servers. Beyond manual inspection it drives your server with real models across many client configurations, scores the results, and runs the same checks in CI.
- A playground that emulates client UIs, tool calls and skills, with Chat, Trace and Raw views, plus a widget emulator that switches between desktop, tablet and mobile and tests locale, CSP permissions, light and dark mode and safe-area insets
- Multi-server chat on frontier models, with up to three models compared side by side and each server's token usage visible
- An OAuth debugger that walks the flow step by step, with conformance checks across protocol versions 2025-03-26, 2025-06-18, 2025-11-25 and 2026-07-28, covering dynamic client registration and pre-registration
- Manual server debugging: tools, resources, resource templates, prompts and elicitation flows run by hand with full JSON-RPC logs
- Evals defined as test cases with expected tool calls, run across LLMs, with accuracy tracked over time so a regression is visible rather than reported
- Skills usable in chat and playground, with local skills read from your filesystem and never leaving the machine
- Workspaces — shared server groups with real-time sync, so a team tests against the same configuration
- A CLI for probing servers, doctor checks, OAuth exercises and listing tools, resources and prompts from a terminal, and an SDK for asserting on capability shapes from your own tests
- CI integration, so conformance, end-to-end tests and evals run on every pull request
Three ways to run it, and they are not equivalent. The hosted app at app.mcpjam.com needs no install but is HTTPS-only — no local stdio servers, tunnelling, skills or tasks. The desktop app for Mac and Windows handles both HTTP and local stdio with no runtime to install. The terminal route, npx @mcpjam/inspector@latest, also covers both and needs Node.js 20 or newer. There is no published Docker image; building one from source is documented, and it should be bound to 127.0.0.1 rather than exposed. Frontier models are available free in chat, or bring your own API key.
One command — npx @mcpjam/inspector@latest