MCP Integration (AI Agent Tools)

dregg exposes its capabilities as MCP (Model Context Protocol) tools over JSON-RPC 2.0 stdio transport. AI assistants (Claude, GPT, etc.) can discover and invoke tools to interact with the federation.

Running the MCP Server

dregg-node mcp --data-dir ~/.dregg --federation-peers "peer1:8420"

This reads JSON-RPC from stdin and writes responses to stdout. Configure your AI assistant to launch this as a tool server.

Available Tools

The MCP server exposes these tools via tools/list:

• authorize -- Authorize an action with a held token
• mint_token -- Create a new root capability token
• attenuate -- Restrict a token's permissions
• delegate -- Hand an attenuated token to another agent
• submit_turn -- Execute a state transition
• post_intent -- Broadcast a capability need
• list_tokens -- Show all held tokens
• node_status -- Federation status and sync state
• seal -- Encrypt data under a capability (X25519-ChaCha20Poly1305)
• unseal -- Decrypt sealed data

Protocol

Implements the MCP subset for tool serving:

• initialize -- Capability negotiation
• notifications/initialized -- Client readiness (no response)
• tools/list -- Enumerate available tools
• tools/call -- Invoke a tool

The JSON-Schema for each tool's input/output is served from the tools/list endpoint at runtime — point your MCP client at a running dregg-node and read the live schema rather than chasing a copy in these docs. End-to-end transcripts and authorisation patterns (read-only cap, scoped writer, federation observer) are covered in docs/agent-authorization.md.