space-standards-mcp

An MCP server (and a thin optional agent) that answers questions about ECSS and NASA space standards with exact, requirement-level citations: every result names the document, issue and requirement ID, or the answer is "not covered". You bring the standards; nothing from them is stored in this repository. It is aimed at launcher and payload-fairing structures first (factors of safety, fracture control, loads, shock, vibroacoustics) and is built to grow to the whole ECSS and NASA set.

Independent project. Not affiliated with or endorsed by the European Space Agency (ESA), the ECSS Secretariat or NASA. The ECSS, ESA and NASA names and the names of their standards belong to their owners and are used here only to identify the documents the software works with.

60-second demo (invented fixtures)

The repository ships three invented demo standards and a demo handbook in tests/fixtures/ (made-up text and numbers, so you can try everything without any real document).

git clone https://github.com/manojknaikade/space-standards-mcp.git && cd space-standards-mcp
python -m venv .venv && source .venv/bin/activate      # Python 3.11+
pip install -e ".[dev]"
pytest -q                                              # fixture tests; real-document tests skip without documents/

# retrieval eval on the demo standards: expected recall@5 1.00 over 4 questions
python eval/run_retrieval_eval.py --docs tests/fixtures --questions eval/demo_questions.yaml

# the MCP server over the demo standards (stdio; an MCP client talks to it)
SPACESTD_DOCS=tests/fixtures python -m spacestd.server

Connect a client (below) to that server and ask, for example, "which factor of safety applies to a metallic fairing in DEMO-E-ST-98C?". The answer comes back with a citation such as DEMO-E-ST-98C Rev.2, 4.1.2a.

Use your own documents

No standards text is distributed here. Download the documents yourself and put them in documents/ (gitignored) or any folder you point SPACESTD_DOCS at:

• ECSS: https://ecss.nl (free registration); Word or PDF.
• NASA: the NASA Technical Standards System, https://standards.nasa.gov. python scripts/download_nasa_standards.py documents/nasa fetches the public ones politely (robots.txt, one request every 2 s); --dry-run only lists them.

Then:

python -m spacestd.catalogue documents      # what was found, and what is loaded
python -m spacestd.health documents         # parse every document, set its tier, write eval/CORPUS_HEALTH.md

documents/corpus.yaml selects what is loaded (doc-ID patterns with *, see ADR-0006); without it only files directly in documents/ load:

load:
  - ECSS-E-ST-32*
  - NASA-STD-5*

File naming, sidecar metadata and formats (PDF preferred; Word files need LibreOffice) are described in documents/README.md. Parsed documents are cached in documents/.spacestd-cache/. The standards remain subject to their publishers' terms of use.

Install without a clone

The package is on PyPI:

uvx space-standards-mcp
pipx install space-standards-mcp

The latest development version installs from git: pipx install git+https://github.com/manojknaikade/space-standards-mcp.

The space-standards-mcp command starts the server (stdio) and reads SPACESTD_DOCS.

Connect an MCP client

Use an absolute path to the Python that has the package installed, and to your documents folder.

With uvx, a client can start the server without a local install, e.g. in Claude Desktop: {"command": "uvx", "args": ["space-standards-mcp"], "env": {"SPACESTD_DOCS": "/abs/path/to/documents"}}.

Claude Code

claude mcp add space-standards --env SPACESTD_DOCS=/abs/path/to/documents -- /abs/path/to/.venv/bin/python -m spacestd.server

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "space-standards": {
      "command": "/abs/path/to/.venv/bin/python",
      "args": ["-m", "spacestd.server"],
      "env": { "SPACESTD_DOCS": "/abs/path/to/documents" }
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "space-standards": {
      "type": "stdio",
      "command": "/abs/path/to/.venv/bin/python",
      "args": ["-m", "spacestd.server"],
      "env": { "SPACESTD_DOCS": "/abs/path/to/documents" }
    }
  }
}

Tools

Tool	What it does
list_documents	Loaded standards and issues with family, type, discipline, parse tier, record counts and tables; filters; can also list documents present but not loaded
get_requirement	Exact record for a document and requirement ID
search_requirements	BM25 keyword search over requirements and table rows, with neighbouring records per hit; reports named standards that are not loaded
get_table	A table (e.g. factors of safety) as structured rows with labels, cell text, parsed value and footnotes
get_references	Standards and clauses a requirement points to
compare_topic	One topic searched separately in each standard, grouped ECSS / NASA, with a coverage rating per standard
get_section_outline	A standard's table of contents; for one clause its record IDs and tables (no requirement text)

Every record carries its kind (requirement, recommendation, permission, guidance, deleted) and a citation.

Optional: hybrid search and the agent

pip install -e ".[embed]" and SPACESTD_EMBED=default add embedding ranking fused with BM25 (opt-in, see eval/RESULTS.md). pip install -e ".[agent]" adds python -m spacestd_agent "<question>" --docs documents, the same tools in a tool-calling loop with a citation gate that checks each draft for invented IDs and numbers not found in the retrieved records (ADR-0005); it needs an Anthropic API key.

Parse tiers

Documents differ in how well they parse, so each carries a tier (ADR-0006):

• verified: checked by hand against the source (requirement counts, IDs, tables).
• checked: passes the automatic health checks (python -m spacestd.health), not reviewed by hand.
• unchecked: parsed, but no check has passed; treat results as leads.

Records, tables and rows carry their tier and page. Numbers taken from tables of documents below verified come with a "check against the source, page N" flag, and the citation gate only accepts such a number when the answer repeats it. Always confirm a number in the source.

Evaluation and its limits

Two evals run on a 57-question set (17 numeric, 17 lookup, 7 classification, 5 compare, 11 not covered). Numbers from eval/RESULTS.md, on the 7 hand-verified standards:

Measure	Value
retrieval recall@5 / @20 (46 scored questions)	0.64 / 0.89 (0.74 at @5 counting neighbouring records)
table recall@5: row hit / reachable (10 questions)	0.50 / 0.85
answers, independent judge, plain client: correct / partial / wrong	52 / 4 / 1
answers, gated agent: correct / partial / wrong	49 / 7 / 1
invented requirement IDs	0
"not covered" questions declined without numbers	11 / 11

Limits: the question set is small and written by one author; the first 21 questions had domain review and the later additions did not; one model family answers and judges; retrieval is keyword-based and misses paraphrases (hybrid search helps only partly); and the figures say nothing about the many documents at tier checked or unchecked. Treat answers as aids for locating requirements, not as authority.

Disclaimer

• No standards text is included. You download the standards yourself.
• Answers are aids, not authority. They are not a substitute for the applicable standard in the issue your project is contracted to, with its tailoring. Verify any requirement, factor or test level against the source.

Project docs

Roadmap · Release plan · Project brief · Architecture · Decision records · Contributing · Security

Licence

The code is licensed under the Apache License 2.0; see also NOTICE and ADR-0007.