clairwave-mcp
An open MCP server that gives AI assistants physically grounded ocean acoustics.
Clairwave runs validated propagation models (Bellhop,
RAM/parabolic equation) on global bathymetry and seasonal sound-speed profiles,
tracks live AIS vessels, and serves 3D hull models for them
(shipshape). This server exposes that
to Claude, ChatGPT, Gemini and any other MCP client — so an assistant reasoning
about the ocean can run the physics instead of guessing.
Every result carries provenance (model, data source, run_id) and an
open_url that opens the exact result in the platform. Simulation results
include the bathymetry, sound-speed profile and bottom parameters that were
used, so a researcher can replicate the run in MATLAB, Python or anything else.
Endpoint (no auth, no key): https://www.clairwave.com/mcp — Streamable HTTP.
Connect
- Claude Code:
claude mcp add --transport http clairwave https://www.clairwave.com/mcp
- Claude.ai / Claude Desktop: Settings → Connectors → Add custom connector → the URL above
- ChatGPT: Plugins → search Clairwave → + (a listed plugin: no URL, no developer mode)
- Any MCP client: point it at the URL; the server is stateless and JSON-response capable
Tools
Typical latency against the live platform: bathymetry 0.5 s, SSP 6 s first time
per 0.1° cell then cached, RAM transmission loss 1–3 s, detection range 1–3 s.
Run locally
Environment: CLAIRWAVE_API, CLAIRWAVE_FLEET, CLAIRWAVE_SITE, MCP_PORT.
Where to find it
- Official MCP Registry:
io.github.clairwave/clairwave (https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.clairwave/clairwave)
- Claude: Settings > Connectors > Add custom connector, URL
https://www.clairwave.com/mcp, no auth.
- ChatGPT: listed in the Plugins directory as "Clairwave — Ocean acoustics & live ships"; search Clairwave and click +.
- Grok (grok.com/connectors > New > Custom): paste the same URL.
- xAI / OpenAI APIs:
{"type": "mcp", "server_url": "https://www.clairwave.com/mcp", "server_label": "clairwave"}.
Place names
Every location tool takes either lat/lon or a place string. Names go through a
maritime gazetteer first (ports resolve to their approaches, straits and seas to a
representative water point; ~120 entries in gazetteer.py), then OpenStreetMap
Nominatim. If the point is on land or shallower than 10 m it is walked seaward until
it is deep enough, and the response's location block reports the original point,
the snap distance and bearing, and the depth used. resolve_place exposes the same
logic directly, with offshore_km to push a point further out.
Example prompts
- "What is the sound speed profile 50 km west of Gibraltar in March, and where is the sonic layer depth?"
- "How far could a 150 Hz, 170 dB source at 20 m depth be detected by a receiver at 100 m near 36N 5.5W, along bearing 090?"
- "Show transmission loss versus range at 200 Hz out to 30 km north of Halifax in winter."
- "What ships are within 15 km of the Strait of Hormuz right now, and how loud is the largest one?"
- "Run a 3D Bellhop volume at 400 Hz around 49.2N 123.3W and give me the link to open it."
Limits and support
License
MIT. Data: AIS via the AISHub peer network (Clairwave contributes receivers);
vessel photos CC-licensed with attribution; bathymetry and SSP sources cited in
each response.