Running in e2e harnesses

mcp-smoke gives an end-to-end test one realistic client. Send a plain-English request, let it go through the real MCP tool path, then check what your infrastructure did: routing, auth, logging, tracing, and whether analytics were generated for a real tool invocation.

A check you can copy

scripts/agent-e2e.sh is the live check this repository runs. It:

1. requires OPENAI_API_KEY and writes it to a temporary env file (.env by default, or the path given as the first argument), then unsets it from the environment so the CLI has to read the file;
2. runs the agent against the bundled test server with the prompt add 41 and 1;
3. asserts that the combined output contains the tool-call trace tool> add {"a":41,"b":1} and the number 42;
4. removes the env file and its log on exit.

The same shape works for your own server: replace --server and the prompt, and assert on the tool> trace and the answer.

In this repository's CI

ci.yml runs on pushes and pull requests to main:

• go test ./...
• a Docker smoke test that builds the image and runs it
• a live agent e2e job that runs the script above only when an OPENAI_API_KEY secret is configured, and is skipped otherwise

The Docker image

The repository's Dockerfile builds both mcp-smoke-agent and the test server into one image. Its entrypoint runs smoke mode against the bundled test server over stdio:

docker build -t mcp-smoke .
docker run --rm mcp-smoke

The bundled test server

cmd/mcp-test-server is a small stdio MCP server for trying the CLI and for testing the CLI itself:

Kind	Names
Tools	echo, add, ping
Prompts	hello, hello-name
Resources	example-resource (mcp://example/resource)