manna-mcp

MANNA

MANNA — MCP Architecture for NOIRLab, NRAO, and Additional 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__*). The one exception is the PyPI distribution, manna-mcp (bare manna is admin-prohibited on PyPI).

Tools

MANNA has four layers. Connections call the standard IVOA interfaces (TAP, SIA, SCS, RegTAP, Sesame). Workflow tools bundle a multi-step task into one call. Result handling returns small results inline and a link plus a fetch recipe for large ones. Archive notes are one file per archive holding its addresses and notes about its quirks, each note with a check that re-verifies it. Every tool below is tagged with the layer it belongs to.

Tool	Protocol	Layer	Description
list_archives	—	Archive notes	List known archives with endpoint URLs and usage notes
describe_table	—	Archive notes	Curated per-table schema facts (missing columns, enum values, spatial index hints)
resolve_target_name	Sesame	Connections	Resolve an object name (e.g. "M87", "Cygnus A") to RA/Dec coordinates
run_adql_query	TAP	Connections · Result handling	Submit sync or async ADQL queries; returns inline or promoted results
get_async_job_status	TAP	Connections	Poll an async job by ID
get_async_job_results	TAP	Connections · Result handling	Return a completed async job's result URL + pyvo fetch recipe (client fetches the data)
abort_async_job	TAP	Connections	Abort a running async job
search_ivoa_registry	RegTAP	Connections	Search the IVOA registry by keyword or service type
describe_ivoa_service	RegTAP	Connections · Result handling	Describe a specific registry resource (columns, capabilities)
search_catalog_by_position	SCS	Connections · Result handling	Simple Cone Search for legacy SCS-only archives
search_images_by_position	SIA 2.0	Connections · Result handling	Search for images by position and waveband (returns access URLs to fetch client-side)
find_observations_of_target	SIA 2.0 / SCS	Workflow tools	One-call workflow tool: resolves a target name or coordinates, auto-selects an archive by service/waveband, then runs the SIA (image) or SCS (catalog) search — chains resolve_target_name + list_archives + search_images_by_position/search_catalog_by_position so the model doesn't have to
count_observations_near_target	TAP	Workflow tools	Count observations/sources near a target in one call (resolve → select archive → COUNT)
survey_archives_for_target	TAP	Workflow tools	Survey which archives hold data for a target, with per-archive counts
preview_table	TAP	Workflow tools · Archive notes	Columns + curated enums/notes + a sample of rows for one table, in one call

Renamed in 0.9.0. Tool names dropped the vo_ prefix for verb-first, descriptive names; clients pinned to the old names must update.

Before 0.9.0	Now
vo_archive_list	list_archives
vo_schema_describe	describe_table
vo_inspect_table	preview_table
vo_target_resolve	resolve_target_name
vo_tap_query	run_adql_query
vo_tap_status	get_async_job_status
vo_tap_results	get_async_job_results
vo_tap_abort	abort_async_job
vo_registry_search	search_ivoa_registry
vo_registry_describe	describe_ivoa_service
vo_cone_search	search_catalog_by_position
vo_sia_search	search_images_by_position
vo_find_observations	find_observations_of_target
vo_count_observations	count_observations_near_target
vo_survey_target	survey_archives_for_target

The recommended LLM workflow for a positional query:

• resolve_target_name — get RA/Dec for a named object
• list_archives — discover the archive and its endpoint
• describe_table — get table-specific quirks before writing ADQL
• describe_ivoa_service — live column introspection
• run_adql_query (mode=async for data reads) — run the query

Install

pip install manna-mcp                # distribution name; the import + CLI are `manna`
manna                                # boots the server on http://localhost:8000
# or run without installing:
uvx manna-mcp

As an MCP server in a client

MANNA speaks stdio with --stdio, and streamable HTTP otherwise. For a stdio client (Claude Desktop, Claude Code, IDE extensions):

claude mcp add manna -- uvx manna-mcp --stdio

or, editing a client config by hand:

{"mcpServers": {"manna": {"command": "uvx", "args": ["manna-mcp", "--stdio"]}}}

The first launch resolves astropy and pyvo, which is a large download; run uvx manna-mcp --stdio once in a terminal before wiring it into a client whose startup timeout is short.

Quickstart

uv sync
uv run pytest --record-mode=none        # 732 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:

Variable	Default	Description
MANNA_PORT	8000	HTTP listen port
MANNA_HOST	0.0.0.0	Bind address
MANNA_LOG_LEVEL	INFO	DEBUG / INFO / WARNING / ERROR
MANNA_TAP_SYNC_TIMEOUT_SECONDS	20.0	Timeout for sync TAP queries
MANNA_ALLOWED_HOSTS	(unset)	Comma-separated hostnames the server may fetch (exact or subdomain match). Unset ⇒ any public host; private/loopback/link-local targets are refused regardless
MANNA_ARCHIVES	(unset)	Comma-separated archive short_names to activate. Unset/empty ⇒ all archives physically present in archives/ except paused ones (nrao ships paused; name it here to re-enable)
MANNA_INLINE_ROW_LIMIT	200	Max rows in an inline result before it's routed to an async job (TAP) or truncated (cone/SIA)
MANNA_INLINE_BYTE_LIMIT	49152	Max bytes in an inline result before the same promotion/truncation applies (48 KiB)
MANNA_REGISTRY_DESCRIBE_BYTE_LIMIT	49152	Above this, describe_ivoa_service degrades from per-column detail to a table catalog (names + descriptions + column counts)

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.
• Paused — an archive can set paused="<reason>" to ship inactive by default without losing its notes. nrao is paused while NRAO rebuilds its TAP service; MANNA_ARCHIVES=datalab,alma,nrao re-enables it.

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

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

Full documentation: https://manna.readthedocs.io — installation, client setup (Claude Code, Claude Desktop, Jupyter AI), a guide to how results and archive notes work, the generated tool reference, and tutorials.

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

Deployment configurations are maintained in a separate repository.