MCP

The Arbiter MCP server lets AI agents interact with the same risk, control, and policy data as the REST API through typed Model Context Protocol tools.

Maturity: Live / Early Access — same operational layer as REST (Integration Surfaces). MCP does not invent a separate verification layer; hashes and optional Seal semantics match Verification.


Endpoint	`https://mcp.arbiter.blockskunk.com/mcp`
Transport	Streamable HTTP (production) or stdio (local)
Auth	`Authorization: Bearer arb_…` (inherits user RBAC — Access Model)

When to use MCP

Use REST when…	Use MCP when…
Scripts, ETL, dashboards	Attaching an LLM to Arbiter
Bulk reads / exports	Runtime tool discovery via `tools/list`
Full Partner Read sync	Interactive agent workflows

Connecting

Cursor — Settings → MCP → Add server:

{
  "url": "https://mcp.arbiter.blockskunk.com/mcp",
  "headers": { "Authorization": "Bearer arb_…" }
}

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "arbiter": {
      "url": "https://mcp.arbiter.blockskunk.com/mcp",
      "headers": {
        "Authorization": "Bearer arb_…"
      }
    }
  }
}

Tools are registered from the same Zod schemas as REST. Field names and enums are byte-identical to GET /api/risks/:uuid.

Browse REST OpenAPI