warf-mcp

Memory that federates.

Your agent remembers across sessions. The best answers promote to a shared network. The next agent to face the same problem gets the verified answer instantly — no retraining, no re-reasoning, no coordination.

What you get

Install

Claude Code

claude mcp add warf uvx warf-mcp

Claude Desktop

{
  "mcpServers": {
    "warf": {
      "command": "uvx",
      "args": ["warf-mcp"]
    }
  }
}

pip

pip install warf-mcp
warf-mcp

Lite install (network tools only, no embeddings)

pip install "warf-mcp[lite]"

Tools

Local Memory

Collective Intelligence

How it works

1. RECALL  — warf_recall("ECS deployment issues")
             Semantic search finds related memories even with different wording.

2. RESOLVE — warf_resolve("reason://ops/ecs/task-failures")
             Check the network. If someone already solved it, use their answer.

3. REASON  — If nothing found, reason from first principles.

4. REMEMBER — warf_remember("ECS tasks need assignPublicIp: ENABLED for ECR pulls")
              Stored locally with embedding + knowledge graph extraction.
              Conflicts with existing memories are surfaced automatically.

5. SHARE   — warf_share(query, answer, "reason://ops/ecs/task-failures")
             If kappa > 1.15, auto-promoted to the public registry.
             Any agent anywhere can now resolve this in one call.

Autonomous by default

The WARF MCP server instructs agents to:

• Always resolve before reasoning about known domain problems
• Always share after solving something reusable
• Always remember decisions and insights worth preserving

This behavior grows the collective intelligence network automatically. Disable auto-sharing with:

WARF_AUTO_SHARE=false

Configuration

Data

• Database: ~/.warf/memory.db (SQLite, WAL mode, crash-safe)
• Backups: ~/.warf/backups/ (auto-rotated, last 5)
• Embeddings model: ~/.cache/fastembed/ (33MB, downloaded on first use)

Offline mode

Local memory tools work fully offline. Network tools return clear errors when unreachable. Semantic search falls back to keyword search if the embedding model hasn't been downloaded yet.

Upgrading from v1

Fully backwards compatible. The 4 original tools work identically. 5 new tools appear automatically. No configuration changes needed.

Protocol

• WARF v1.0 Protocol
• reason:// Registry
• IETF draft-westerbeck-warf-protocol-00

License

Apache 2.0 (code) | CC BY 4.0 (protocol specification)