Runnable examples for mcp-devtools.
Each example is a self-contained directory you can cd into and run in under a minute.
| Directory | What it shows |
|---|---|
01-proxy-basic |
The 60-second tour. Spin up a fake MCP echo-server, proxy it, see frames stream into the inspector. |
02-record-and-share |
Capture a session to .mcptrace and reopen it as a shareable bug-report artifact. |
03-embed-in-server |
Bake the inspector into your own MCP server using the mcp-devtools/embed API. |
04-http-transport-with-cost |
Proxy a Streamable-HTTP + SSE MCP server and watch per-call USD cost attribution roll up. |
05-record-and-diff |
Record a baseline session, record a session against a regressed server, and diff the two .mcptrace files to catch protocol-level breakage. |
06-embed-mode-walkthrough |
Use devtools.wrap(transport) to bake the inspector into your own MCP server in-process — same observability as the proxy, no child process, no stdio dance. |
07-doctor-ci |
Wire mcp-devtools doctor --json into a GitHub Actions workflow that fails CI on protocol-compliance regressions (missing tool fields, malformed envelopes, dropped tools/list). |
08-replay-snapshot-ci |
Commit a recorded .mcptrace as a test fixture and replay it via serve --replay so client snapshot tests run against a deterministic, vendor-free MCP backend in CI. |
09-cost-budget-alert |
Wire mcp-devtools summary --model --json into GitHub Actions so a recorded session's estimated USD cost gates the build — catch prompt-growth and tool-call creep regressions before they ship. |
git clone https://github.com/adityachilka1/mcp-devtools-examples
cd mcp-devtools-examples/01-proxy-basic
pnpm install # or npm / bun
pnpm startEach example has its own README.md with the full walkthrough.
- The wire-level model of MCP — what the JSON-RPC envelopes actually look like for
tools/list,tools/call, notifications. - How to integrate the inspector into your existing dev loop with zero changes to your server code.
- How to use recorded sessions as bug-report artifacts that don't require screenshots.
- How to think about protocol-level testing for an LLM agent stack.
mcp-devtools— the inspector itself.skillforge— CLI + registry for Claude Skills.agentbench— snapshot tests for AI agent traces.
MIT © 2026 Aditya Chilka.