astrodynamics-mcp

A Model Context Protocol server that gives any MCP-capable LLM client (Claude Code, Cursor, ChatGPT desktop, custom agents) authoritative astrodynamics tools: TLE/SGP4 propagation, Lambert solving, ground-station access, time-scale and coordinate-frame conversions, porkchop scans, B-plane targeting, satellite metadata, and — with optional extras — full NASA GMAT mission execution ([gmat]), NASA SPICE / NAIF kernel queries ([spice]), and trajectory visualisation ([viz]).

Why

LLMs reason well about astrodynamics concepts but cannot do the numerical work — they cannot propagate orbits, solve Lambert problems, or query SPICE ephemerides. astrodynamics-mcp lets you plug authoritative tools into any MCP-capable client so the LLM calls vetted upstream libraries instead of fabricating numbers. Every result carries explicit units; every tool description tunes against an Inspect AI eval suite that measures whether the LLM picks the right tool and binds the right arguments.

Tools

† Credentialed source. Pass credentials as environment variables for the stdio transport, or in the session-init _meta block for HTTP — see Credentials. A tool called without its credential returns a typed CredentialRequiredError, never a silent failure.

GMAT tools (optional [gmat] extra)

Install the [gmat] extra and have a local NASA GMAT install, and five more tools register for driving real GMAT missions (they stay hidden otherwise):

SPICE tools (optional [spice] extra)

Install the [spice] extra and seven more tools register, backed by NASA NAIF's CSPICE through spiceypy (they stay hidden otherwise). They furnish kernels into a process-global pool and query whatever the pool holds:

The kernel model, the NAIF furnish-from-URL allowlist, and the process-global pool's trust boundary are covered on the SPICE integration page.

Visualisation tools (optional [viz] extra)

Install the [viz] extra and four more tools register, backed by matplotlib (static PNG plots) and the gmat-czml sibling (CZML export) — they stay hidden otherwise. Each returns its picture as an attachment alongside a numeric summary, so a text-only client still gets the answer:

The attachment model — additive PNG ImageContent / CZML EmbeddedResource beside the structured summary — and which clients render each kind are covered on the Visualisation page.

Full input / output JSON schemas live on the Tool reference page of the docs site.

Quick start

Install:

uv tool install astrodynamics-mcp            # or: pipx install astrodynamics-mcp
uv tool install "astrodynamics-mcp[gmat]"    # adds the GMAT mission tools (needs a local GMAT install)
uv tool install "astrodynamics-mcp[spice]"   # adds the SPICE tools (pulls spiceypy / bundled CSPICE)
uv tool install "astrodynamics-mcp[viz]"     # adds the visualisation tools (pulls matplotlib / gmat-czml)

Claude Code

Add to your Claude Code MCP settings:

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

Restart Claude Code. In a chat:

You: Compute the Hohmann Δv from a 250 km circular LEO to GEO.

(The model calls lambert_solve with the Hohmann geometry and answers ≈ 3.91 km/s, citing the tool output — not the LLM's own weights.)

Cursor

~/.cursor/mcp.json (or workspace-level .cursor/mcp.json):

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

Restart Cursor. The tools appear under the astrodynamics server group.

See Pick a client in the docs for ChatGPT desktop, a raw Python MCP smoke client, and the Streamable HTTP transport for remote agents.

Supported clients

What this is not

• Not a general-purpose astrodynamics framework. Wraps vetted upstream libraries; does not re-implement propagators, integrators, or coordinate systems.

• Not an agent framework. Exposes MCP tools; LangGraph, AutoGen, CrewAI, and the LLM clients themselves consume them.

• Not an ML / inference server. Tools that need their own ML models (maneuver detection, neural propagators) belong in separate MCP servers — kept modular for dependency isolation.

• Not a SaaS. Runs locally or in your own infrastructure. No hosted multi-tenant deployment.

• Not a web UI. Tool consumption is via MCP clients; no browser frontend, no desktop app, no notebook widget.

For direct (non-MCP) Python use of the same surfaces, reach for the upstream libraries: sgp4, lamberthub, skyfield, astropy, interplanetary-porkchop, spiceypy.

Built on

The official Anthropic modelcontextprotocol/python-sdk (MIT). The bundled FastMCP server class is the server primitive; stdio + Streamable HTTP transports are first-class.

Docs and links

• Docs site: astro-tools.github.io/astrodynamics-mcp — getting started, per-client setup, tool reference, recipes, visualisation, data sources, eval suite, FAQ.

• Issue tracker: astro-tools/astrodynamics-mcp/issues

• Discussions: orgs/astro-tools/discussions — usage help and open-ended questions.

• Eval suite: eval/README.md — the regression contract on tool-description quality.

Privacy

astrodynamics-mcp runs entirely on your own machine and collects nothing — no telemetry, no analytics, no accounts. The only data that leaves your machine is the query parameters a tool sends to the data source it wraps (CelesTrak / JPL Horizons / IERS with no auth, and — only if you configure their credentials — Space-Track and ESA DISCOSweb). Credentials are read from local environment variables or the session _meta block and are sent only to their own service over HTTPS. See the Privacy page for the full breakdown.

License

MIT — see LICENSE.