miqa-mcp 0.1.0

MIQA MCP Server

MIQA MCP

A read-only MCP server built with FastMCP. It exposes MIQA test-reporting data to Claude Code (and any MCP client) over stdio. It is a thin MCP surface on top of the miqatools library: the tools call miqatools' read-only reporting wrappers, which resolve credentials and talk to the MIQA v2 REST API.

The packaging shape — a console script that runs a stdio main(), configured via environment variables — is what makes it registerable with claude mcp add.

Tools exposed (all READ-ONLY; none mutate MIQA state):

• find_runs_by_version(version_name: str, limit: int = 100) -> list — find Test Chain Runs for a version name (newest-first; empty list if none match).
• get_test_run(run_id: int, sample_ds_id: int | None = None) -> dict — JSON details for a single Test Chain Run (optionally narrowed to one sample).
• get_test_run_results(run_id: int, sample_ds_id: int | None = None) — the assertion/result summary table (JSON or CSV depending on the server).
• get_test_run_report(run_id: int, sample_ds_id: int | None = None, format="json") — the sample-centric structured report (feature-flag gated on the server).

Built on the standalone fastmcp package (jlowin). Requires Python >= 3.10.

Configuration

The server reads its MIQA credentials from environment variables (resolved by the miqatools client):

• MIQA_SERVER_URL — MIQA server hostname (e.g. yourco.miqa.io).
• MIQA_API_KEY — MIQA API key.

Claude Code injects these into the server subprocess via the --env block (see "Register with Claude Code" below).

Prerequisite: install miqatools first (not yet on PyPI)

This server depends on miqatools>=2.0.0rc0, which is not yet published to PyPI. Until it is, install miqatools from its local checkout first (editable), then install this server:

pip install -e ../magna-cli   # the miqatools library
pip install -e .              # this server

Once miqatools is published, pip install . (or uvx miqa-mcp) will resolve it from PyPI automatically and this extra step goes away.

---

1. Install options

Because miqatools is not yet on PyPI (see "Prerequisite" above), install it locally first, then install this server. A plain virtualenv with editable installs is best while miqatools is unpublished:

python -m venv .venv
.venv/bin/pip install -e ../magna-cli   # miqatools (the library dependency)
.venv/bin/pip install -e .              # this server

The console script then lives at .venv/bin/miqa-mcp. It starts and waits silently on stdin — that is correct: it is a stdio MCP server waiting for a client to speak the protocol. Press Ctrl-C to stop; you normally let Claude Code launch it rather than running it by hand.

Eventual published form

Once both miqatools and this server are published to PyPI, the local-editable dance collapses to a normal install (dependencies resolve automatically):

uvx miqa-mcp        # ephemeral
pipx install miqa-mcp

---

2. Register with Claude Code

claude mcp add <name> [flags] -- <command> [args...] registers a stdio server. Claude Code launches <command> as a subprocess and speaks MCP over its stdin/stdout. The -- separates Claude's own flags from the subprocess command line.

Since miqatools is unpublished, install into a venv (see above) so miqa-mcp is on the venv's PATH, then register that console script — passing MIQA credentials via --env:

claude mcp add miqa-mcp \
  --env MIQA_SERVER_URL=yourco.miqa.io \
  --env MIQA_API_KEY=your-key-here \
  -- /absolute/path/to/.venv/bin/miqa-mcp

If miqa-mcp is already on your PATH (activated venv or pipx install), the command is just the script name:

claude mcp add miqa-mcp --env MIQA_SERVER_URL=yourco.miqa.io --env MIQA_API_KEY=your-key-here -- miqa-mcp

Passing configuration via env

A stdio server cannot receive HTTP headers, so its configuration is passed as environment variables, which Claude Code injects into the subprocess. Use --env KEY=value (short form -e KEY=value), and put it before the -- so it is read as one of Claude's own flags.

This server reads MIQA_SERVER_URL and MIQA_API_KEY from its environment (the miqatools client resolves them), so both must be passed via --env:

claude mcp add miqa-mcp --env MIQA_SERVER_URL=yourco.miqa.io --env MIQA_API_KEY=your-key-here -- /absolute/path/to/.venv/bin/miqa-mcp

Repeat --env for each variable. This env-block mechanism is the stdio equivalent of the --header flag that HTTP-transport servers use; stdio servers get their config through env, not headers.

By default the server is registered at local scope (this project only). Add --scope user to make it available across all your projects.

---

3. Verify

List registered servers — miqa-mcp should appear, and Claude Code will have started it and confirmed the connection:

claude mcp list

You can also inspect just this one:

claude mcp get miqa-mcp

Inside a Claude Code session, the tools surface as find_runs_by_version, get_test_run, get_test_run_results, and get_test_run_report. With valid MIQA_SERVER_URL / MIQA_API_KEY in the env, ask the session to find runs for a version name and it will call the MIQA API through the server.

---

4. Remove

claude mcp remove miqa-mcp

If you installed with pipx and want to uninstall the package too:

pipx uninstall miqa-mcp

---

5. Run the tests

The test suite asserts the four tools register on the FastMCP instance and that each delegates to the corresponding miqatools wrapper (the wrappers are mocked, so no network is touched). Install miqatools first, then this package with its dev extra:

.venv/bin/pip install -e ../magna-cli
.venv/bin/pip install -e '.[dev]'
.venv/bin/pytest

---

6. How this fits together

This server is a thin, read-only MCP surface; all MIQA client logic lives in the miqatools library. The reporting wrappers (find_tcrs_by_version, get_tcr_summary, get_tcr_report) and the existing get_tcr_info_json helper stay in miqatools as client code; this repo only adds the FastMCP tools and the stdio entry point. The packaging shape — console script -> stdio main() -> env-based config (MIQA_SERVER_URL, MIQA_API_KEY) — is what claude mcp add launches.