MANNA

MANNA — MCP Architecture for NOIRLab and NRAO Archives.

An MCP server exposing IVOA-compliant astronomical archives (NOIRLab Astro Data Lab, NRAO/ALMA, CADC, ESO, Gaia, …) to LLM clients.

Naming: MANNA in prose (it's an acronym); lowercase manna for every identifier — the Python package, python -m manna, the manna:dev image tag, and the MCP client alias (mcp__manna__*).

Tools

The recommended LLM workflow for a positional query:

1. vo_target_resolve — get RA/Dec for a named object

2. vo_archive_list — discover the archive and its endpoint

3. vo_schema_describe — get table-specific quirks before writing ADQL

4. vo_registry_describe — live column introspection

5. vo_tap_query (mode=async for data reads) — run the query

Related MCP server: aqc-mcp

Quickstart

uv sync
uv run pytest --record-mode=none        # 657 tests, offline replay
uv run python -m manna                  # server on http://localhost:8000

Smoke test with MCP Inspector:

npx -y @modelcontextprotocol/inspector --cli http://localhost:8000/mcp --method tools/list

Development

uv sync                        # install runtime + dev deps
uv run pre-commit install      # enable git pre-commit hooks (once per clone)

uv run ruff check .            # lint
uv run ruff format .           # format
uv run pyright                 # type check (src/, basic mode)
uv run pre-commit run --all-files   # run every hook over the whole tree

Pre-commit runs ruff (lint + format), file-hygiene checks, and pyright on each commit; the full test suite runs in CI, not at commit time.

Branch flow (see CLAUDE.md for detail): feature branches <initials>/<name> branch off dev and PR into dev; dev is promoted to main via PR. main is protected — it only advances through PRs with passing CI.

Configuration

All settings are optional — defaults work for local dev. Set via environment variables prefixed MANNA_ or in a .env file:

See .env.example for a template.

Docker

docker build -t manna:dev .
docker run -p 8000:8000 manna:dev

Forking for a specific deployment

This repo is the multi-archive base. Each archive is one self-contained file — its endpoints, usage notes, and per-table schemas all live in src/manna/archives/<short_name>.py. Shape which archives make curated claims two ways:

• Physical — delete the unwanted src/manna/archives/<short_name>.py files. Discovery picks up whatever remains; no other file needs touching.

• Runtime — set MANNA_ARCHIVES=datalab,alma (comma-separated short_names) to narrow a shared image without deleting files. Unset/empty ⇒ every archive active.

A dropped or deselected archive loses only the server's curated claims about it — never its reachability. It's still reachable via vo_registry_search.

Refreshing recorded cassettes

Tests replay archive HTTP traffic from YAML cassettes in tests/<area>/cassettes/. To refresh a stale cassette:

# requires network access to the archive endpoint
rm tests/<area>/cassettes/<test_module>/<test_name>.yaml
uv run pytest tests/<area>/<test_module>.py::<test_name> --record-mode=once

Inspect the cassette diff before committing — large changes in the VOTable namespace URI or response headers may indicate an upstream breaking change.

Docs

• docs/archives-spec.md — how per-archive knowledge modules work, and how to author one

Deployment configurations are maintained in a separate repository.