m3 MCP server

M3 tests MCP servers and the agents that use them. It runs Python and pytest tests, captures MCP evidence, and can save runs for inspection and comparison.

What a test looks like

A test gives an agent a prompt and your MCP server. M3 records every MCP call the agent makes, so the test can check the calls, not only the answer.

import sys

import pytest
from m3 import StdioServer, expect

pytestmark = pytest.mark.m3(agents=[{"harness": "claude-code", "models": ["claude-sonnet-5-5"]}])

registry = StdioServer(name="registry", command=sys.executable, args=("registry_server.py",))


def test_adds_current_tailwind(agent):
    result = agent.run("Add Tailwind to this project.", server=registry)
    expect(result).to_have_tool_call("get_latest_version", arguments={"package": "tailwindcss"})

If the agent answers from memory and never calls the server, the test fails. The fix is usually the tool description, not the test:

 @mcp.tool(
-    description="Get package info",
+    description="Current published version of a package. Call before adding "
+    "or upgrading a dependency; your training data is out of date.",
 )
 def get_latest_version(package: str) -> str:

def test_revenue_by_month(agent):
    result = agent.run("Show me revenue by month.", server=db)
    expect(result).to_have_tool_calls(["describe_table", "query"])

 @mcp.tool(
-    description="Run a SQL query",
+    description="Run a read-only SQL query. Call describe_table first; "
+    "column names are specific to this database.",
 )
 def query(sql: str) -> list[list]:

def test_last_weeks_orders(agent):
    result = agent.run("Get last week's orders from our API.", server=api)
    expect(result).to_have_tool_calls(["find_endpoint", "call_endpoint"])

 @mcp.tool(
-    description="Search the API spec",
+    description="Look up the real path before calling any endpoint. "
+    "Paths are specific to this API; don't guess them.",
 )
 def find_endpoint(query: str) -> str:

Put ANTHROPIC_API_KEY in .env at the project root, then run the tests:

m3 test -- tests

Run the same tests with another agent, or several times to get a pass rate. --trials repeats tests that take the agent fixture, like the ones above; other tests run once.

m3 test --harness codex=gpt-5.6-sol -- tests
m3 test --trials 5 -- tests

Install and start

Install the CLI with uv:

uv tool install sf-m3-cli

On macOS or Linux, the shell installer is an alternative:

curl -LsSf https://m3.sineframe.com/install.sh | sh

Then initialize your project:

m3 init
m3 setup
m3 doctor

Then follow the first MCP test. It uses a local server and needs no model credentials.

Agent skill

m3 init and m3 setup install the testing-with-m3 agent skill for the installed M3 release. To install it yourself:

npx --yes skills@1.7.0 add sineframe/m3#vVERSION --skill testing-with-m3 --agent universal claude-code -y

VERSION is the output of m3 --version. See the agent skill guide.

Submitting a server to the Claude, OpenAI or Muse directory? Install the preparing-directory-submissions skill the same way, with --skill preparing-directory-submissions. It runs m3 lint --directory, does the checks that need judgment and writes one readiness report; see Prepare a directory submission.

Documentation

• Documentation
• Test a local server
• Test agent behavior
• Configure credentials
• Connect an ACP-compatible agent
• Compare agent harnesses and versions
• Pin and inspect harness runtimes
• Handle MCP elicitation
• CLI reference
• Python reference
• Blog
• Contributing

The SDK and CLI use separate environments. The CLI runs pytest in the project environment and saves history; tests import the SDK.

Development

Read Architecture, then use the package README for the area you are changing. Documentation changes follow the documentation writing guide.

just setup
just test-all
just lint
just format-check

M3 is licensed under Apache-2.0.