MRD — machine-readable datasheets for spacecraft components

Typed, datasheet-sourced specifications for 2,486 flight hardware products across 24 categories — reaction wheels, star trackers, propulsion, radios, antennas, structures, ground stations and more — with 19,011 values that each cite the datasheet page they came from.

Built for software and AI assistants, not only for reading. Ask in plain language, filter by engineering requirements, or download the whole thing.

"Find a reaction wheel over 0.2 N·m·s that runs on a 12 V bus, under 2 kg"

A  meets every requirement          VRW-B-02 · Vectronic · 0.2 N·m·s · 9–36 V · 1000 g
B  fails a named requirement        RW-0.06 · momentum 0.06 N·m·s — need ≥ 0.2
C  could fit, data not published    NanoTorque GSW-600 · missing: supply voltage

Three rules make the data usable by a machine:

Nothing is guessed. A missing property means the vendor does not publish it, and the item is reported as "data missing" rather than silently dropped or filled with a typical value.
Conditions are kept apart. Idle, steady-state and peak power are separate records with their measurement condition attached — never averaged into one number.
Every value is traceable. Source file and page, plus a confidence level: L0 machine-extracted, L1 verified by a human against the vendor datasheet.

Use it from an AI assistant (MCP)

Hosted — no install. Add this as a custom connector in Claude, ChatGPT (developer mode) or any MCP client:

https://newspacemarket.com/mrd-connect/mcp

Local — for editors that speak stdio (Cursor, Windsurf, VS Code, Claude Desktop). Add to your MCP config:

{
  "mcpServers": {
    "mrd": {
      "command": "npx",
      "args": ["-y", "@newspacemarket/mrd-mcp"]
    }
  }
}

The package is a thin bridge: it speaks stdio to your editor and forwards to the hosted endpoint, so new tools appear without upgrading anything.

Three tools arrive in the assistant:

tool	what it does
`match_components`	reverse search — requirements in, candidates grouped A / B / C with per-parameter verdicts
`get_component_specs`	the full typed layer of one category, for your own comparison or trade study
`build_budget`	mass and steady-state power of a stack, with bus-voltage checks and an honest list of what could not be counted

Where it is listed

Official MCP Registry — io.github.newspacemarket-com/mrd
Smithery — info-zrja/newspace-mrd, one-click install into Claude, Cursor, VS Code and other clients
Dataset mirror — Hugging Face

Use it from code

No authentication for reading. An optional free key raises rate limits five-fold — request one at /mrd-connect/.

curl "https://newspacemarket.com/mrd-connect/match?category=star-trackers&accuracy_max=10&bus_v=12"

examples/python/select_wheels.py — reverse search, printing why each candidate passed or failed
examples/python/budget.py — assemble a stack and budget it through the MCP tools
examples/js/match.mjs — same search from Node, no dependencies
examples/curl.md — every endpoint as a one-liner

OpenAPI 3.1: openapi.json, served live at https://newspacemarket.com/mrd-connect/openapi.json. Import it as a ChatGPT Custom GPT Action, a LangChain / CrewAI / n8n toolkit, or convert it to Gemini function declarations.

Download the data

data/ holds a dated snapshot, regenerated as the layer grows:

file	contents
`mrd-full.json`	everything — products, values, property dictionary
`mrd-values.csv`	one row per value: product, property, number, unit, bound, condition, source, confidence
`mrd-products.csv`	one row per product: name, manufacturer, category, URL
`mrd-properties.csv`	the property dictionary — what each key means and its canonical unit

Live data is always newer than the snapshot; use get_component_specs or /mrd-connect/specs?category=… when freshness matters.

How the data is built

Values are extracted from vendor datasheets (text and, for scanned documents, page images), from vendor product pages, and from product descriptions — the source of every record says which. Units are converted by code, not by a model: kg to g, dBm to W, psi to bar, tf to mN. A validation pass rejects a figure whose unit does not match the property, and an outlier scan flags values that sit far from the category median for human review.

Coverage is not exhaustive. Where a vendor publishes nothing machine-readable, the product simply is not here — and where a required field is missing, the API says so by name instead of guessing.

Licence

Code (the MCP bridge and examples): MIT — see LICENSE.
Data in data/: CC BY 4.0, attribution "NewSpace Market" — see data/LICENSE-DATA.

Specifications are extracted from publicly published vendor materials and are provided as-is. Verify against the vendor datasheet before committing to a design.