.-----.
  |  🔴  |
  |  🟡  |
  |  🟢  |
  '-----'
   greenlight

greenlight

Listed in awesome-mcp-devtools.

See what your MCP server is actually doing.

Real trace, from an actual recorded session (examples/demo-session.jsonl), not staged text. Green for a clean success, yellow for a slow-but-fine call, red for a tool that actually failed.

A transparent stdio proxy for the Model Context Protocol. Point it at your real server command instead of running that command directly, and it relays every byte exactly as before, while recording every JSON-RPC message to a structured log you can watch live or replay.

Right now, if an MCP integration isn't working, you're debugging blind: no visibility into what got sent, what came back, or why a call failed. Greenlight exists to fix that.

Why not MCP Inspector? Different job. Inspector is a UI you drive yourself to manually test a server in isolation. Greenlight sits transparently in the path of your real client (Claude Desktop, Claude Code, whatever's actually running) and records what really happened, not what you tried by hand. It also writes a durable log (greenlight stats exits non-zero on failure, so it runs in CI) instead of a live session you have to be watching.

Install

pip install greenlight-mcp

Or from source:

pip install -e .

Try it right now, no server of your own required

greenlight run -- npx -y @modelcontextprotocol/server-everything stdio

That's the official MCP reference server -- public, free, no config. In another terminal:

greenlight tail -f

Watch real tool calls come through live. Needs Node.js for the npx part; nothing else.

Use it for real

Not sure how to wrap your actual server command? Check first:

greenlight wrap

Finds your real MCP client config (Claude Desktop, Claude Code, Cursor, or Windsurf) and prints exactly what each server entry would look like rewritten to run through Greenlight. Read-only -- it never touches the file, just shows you what to paste in yourself.

Wherever you'd normally configure a server command, wrap it the same way:

greenlight run -- npx -y @some/mcp-server

instead of

npx -y @some/mcp-server

For a remote Streamable HTTP server, proxy it instead of spawning a process:

greenlight run --http http://127.0.0.1:9000/mcp

Greenlight prints the local URL to point your client at (the same path as the target, just on 127.0.0.1:8808 -- see the printed message, which includes the exact path). Same session log, same tail/stats downstream, regardless of which transport produced it.

Every message that passes through gets logged to ./sessions/. Watch it:

greenlight tail                 # replay the most recent session
greenlight tail -f              # follow a session that's still running
greenlight tail path/to/log.jsonl

Trace output is colorized by status: green for a clean success, yellow for a slow-but-fine call, red for anything that actually failed -- including MCP tool-level failures (result.isError), not just transport-level JSON-RPC errors, which are a different thing and easy to miss if you only check for the obvious one. See notes/day1.md for why that distinction mattered enough to write a whole note about it.

Or skip watching it and just get the summary:

greenlight stats                # message counts, latency, pass/fail
greenlight stats --json         # same thing, machine-readable

stats exits non-zero if anything failed -- transport error or tool error -- so it works as a CI check, not just an interactive summary:

greenlight run -- npx -y @some/mcp-server &
# ... drive a real session against it ...
greenlight stats || exit 1

Let an agent read the trace directly

pip install greenlight-mcp[serve]
greenlight serve

Exposes this project's own session data as an MCP server. Add it to an agentic MCP client's own config (Claude Code, Cursor, anything MCP-capable) and it can call get_failures or get_trace on your last session directly -- the real error messages, not a paste of terminal output. Four tools: list_sessions, get_session_stats, get_failures, get_trace. Optional dependency -- the base install stays just rich.

How it works

greenlight run spawns your real server as a subprocess and sits between it and the real MCP client, relaying stdin/stdout on two threads. Every line is parsed as JSON-RPC, correlated by request id (so a response knows its own method name and latency), and written to a JSONL file. The one rule the whole thing depends on: nothing but the child process's actual bytes ever reaches Greenlight's own stdout -- logging and UI output only ever go to stderr or to disk. A single stray print to stdout would corrupt the protocol stream the real client is parsing.

Status

• greenlight run -- transparent proxy, validated end-to-end against a real MCP server (not a mock)
• greenlight tail -- live trace viewer, both static replay and genuine live-follow (verified separately, not assumed)
• Windows PATH resolution for npx-style commands, validated against a real third-party npx-launched server (the official MCP reference server), not just the Python fixture
• Published to PyPI -- pip install greenlight-mcp
• greenlight stats -- summary + CI-usable exit code (non-zero on any failure, transport or tool-level)
• Streamable HTTP transport (greenlight run --http <url>), validated end-to-end against a real HTTP+SSE server, not just stdio
• Property-based tested (Hypothesis) against arbitrary chunking of the SSE stream -- found and fixed a real Unicode line-boundary bug (str.splitlines() treats more than \n/\r as a line break) that hand-written test cases hadn't caught
• Fuzzed the JSON-RPC classifier too -- found and fixed a crash on any valid-but-non-object JSON (null, 42, [1,2,3]), which json.loads() accepts but a dict-shaped assumption didn't handle
• Failed tool calls show their real error text, not just a red flag -- tail and stats both surface the actual message MCP gave for the failure, pulled from the same result.content a client would show a user
• greenlight wrap -- finds your real Claude Desktop, Claude Code, Cursor, or Windsurf config and shows exactly how to point it at Greenlight, read-only
• greenlight serve -- exposes session/trace data as an MCP server itself, so an agentic client can query real failure data directly instead of a human relaying terminal output to it (optional dep, base install unaffected -- verified against a clean venv)

Notes

notes/ is a running engineering log, not a cleaned-up retrospective -- what broke, how it was found, why the fix is what it is.