@capybaralabs/hive-mcp

MCP server that exposes a HIVE queen as native tools to any MCP-aware LLM host: Claude Desktop, Claude Code, Cursor, Continue, Goose, OpenClaw, and others.

HIVE is a decentralized, verifiable knowledge base for LLMs — peer-to-peer extraction of signed knowledge fragments, embedded vector search, optional LLM synthesis. This package is the bridge between any MCP-aware client and a running HIVE queen.

Reference: USE-CASES.md → case 14 in the HIVE repo.

Install

Use directly (no install) for the canonical setup in a Claude Desktop / Cursor MCP config:

npx @capybaralabs/hive-mcp

Or install globally:

npm install -g @capybaralabs/hive-mcp
hive-mcp

Configuration

Two environment variables:

| Variable | Required | Default | Purpose | |---|---|---|---| | HIVE_QUEEN_URL | no | http://localhost:8090 | Base URL of the queen's HTTP API | | HIVE_API_KEY | when queen auth is enabled (v0.9+) | (none) | Bearer token sent as Authorization header |

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "hive": {
      "command": "npx",
      "args": ["-y", "@capybaralabs/hive-mcp"],
      "env": {
        "HIVE_QUEEN_URL": "https://your-queen.example.com"
      }
    }
  }
}

Claude Code

claude mcp add hive npx -y @capybaralabs/hive-mcp \
  -e HIVE_QUEEN_URL=https://your-queen.example.com

Cursor

~/.cursor/mcp.json (same JSON shape as Claude Desktop).

OpenClaw

Discovered automatically by mcporter from npm — see docs.openclaw.ai/cli/mcp.

Tools

| Tool | Purpose | |---|---| | hive_query | Retrieve signed knowledge fragments matching a free-text query. Returns top-k fragments with score, snippet, URL, signing bee. No LLM synthesis — the calling LLM (Claude/Cursor) composes the answer from the fragments. | | hive_list_sources | List the bees registered on this queen and the sources each one declares. Useful for discovery. |

Why no LLM synthesis in the MCP path?

HIVE's product is the information (signed fragments with provenance), not the prose. When called from an MCP host, the host's own LLM (Claude Sonnet, Cursor's model, etc.) is invariably stronger than any model the queen-side might call — running fragments through a weaker LLM and then a stronger one is redundant and loses fidelity. So hive_query returns raw fragments and lets the host LLM synthesise.

The queen's /api/query endpoint still supports LLM synthesis (use_llm: true) — it's used by the queen's own web UI, by non-MCP integrations, and by case 06 (Local AI · fully-offline stack). Just not from MCP.

Develop

git clone https://github.com/capybarist/hive-mcp
cd hive-mcp
npm install
npm run dev   # runs from src/ with tsx (no build step)
npm run build # emits dist/

Point at a local HIVE queen running on http://localhost:8090 (default), or set HIVE_QUEEN_URL to a remote one.

Transport

stdio only in v0.1 — the universal denominator across MCP hosts. SSE / HTTP-streamable transports are tracked for v0.2 (needed for remote multi-user queens).

License

BUSL-1.1 — same as HIVE.