Marine

Model Context Protocol server for the Ocean Intelligence System. It serves the Blue Life Commons corpus to AI agents (Claude Desktop, Claude Code, Cursor, any MCP client) under two non-negotiable rules:

1. Detail-serving is review-gated. An artifact's prose body is returned as fact only when it is review-approved (status ∈ {approved, published}, and for science-sensitive types review.science = approved). Otherwise the tool returns a typed refusal (servable: false) carrying the artifact's status — never the unreviewed body.
2. Every response carries its sources. Tool results are wrapped in { data, sources, status, attribution }. The IS never launders an unsourced claim.

This is the IS-engine layer of the Ocean Intelligence System triad:

Install & run

npm install
npm run build
BLC_PATH=/path/to/blue-life-commons npm start

Dev (no build step):

BLC_PATH=/path/to/blue-life-commons npm run dev

Configure in Claude Desktop / Claude Code

{
  "mcpServers": {
    "marine": {
      "command": "node",
      "args": ["/abs/path/to/marine-mcp/dist/index.js"],
      "env": { "BLC_PATH": "/abs/path/to/blue-life-commons" }
    }
  }
}

Tools

Why a separate repo (not inside ocean-intelligence-system)

The OS repo previously embedded an MCP server that hardcoded a sibling path, ran a stale protocol version, served needs-expert-review content as fact, and dropped sources from responses. marine-mcp exists to do that job correctly and to keep the IS-engine ("serve the corpus to any agent") cleanly separable from the OS-runtime ("render a dashboard").

Test

npm test        # node:test over the trust-gate logic
npm run typecheck

License

MIT (code). Served content is © its contributors under CC-BY-4.0 per the commons.

Built on SIP · Blue Life Commons — an initiative of Starlight Intelligence Systems.