@guyghost/swarm-dao-mcp

Swarm DAO governance as an MCP (Model Context Protocol) stdio server. Exposes the full Swarm DAO toolset — propose, deliberate, control, execute, ship, rollback, audit — as MCP tools that any MCP-compatible host can call.

Integrating a host? Read docs/MCP_INTEGRATION.md — the canonical workflow, spawn patterns, and operating rules shared by all adapters (Copilot, Claude Code, Codex, and this server).

Install

npm install @guyghost/swarm-dao-mcp

Run

# Boots the stdio MCP server, using $DAO_ROOT (or cwd) as the DAO root.
swarm-dao-mcp

Configure in a host

Any MCP host can register this server. Example (VS Code / Claude / generic):

{
  "mcpServers": {
    "swarm-dao": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@guyghost/swarm-dao-mcp"],
      "env": { "DAO_ROOT": "." }
    }
  }
}

For ready-made host packages (native config + instructions + slash commands), see the sibling adapters:

• @guyghost/swarm-dao-copilot-adapter
• @guyghost/swarm-dao-claude-adapter
• @guyghost/swarm-dao-codex-adapter

Tools

32 tools covering the full governance lifecycle:

| Group | Tools | | --- | --- | | Onboarding | dao_help, dao_setup | | Proposals | dao_propose, dao_update_proposal, dao_propose_amendment, dao_list | | Deliberation | dao_deliberate, dao_record_outputs, dao_roundtable | | Delivery | dao_plan, dao_artefacts, dao_dry_run, dao_control, dao_execute, dao_ship, dao_rollback | | Tracking | dao_dashboard, dao_audit, dao_rate, dao_attention | | Agents | dao_agents | | GitHub | dao_config_github, dao_github_create_branch, dao_github_open_pr | | Workflow runs | dao_graph_status, dao_graph_submit, dao_product_status, dao_product_submit, dao_improve_status, dao_improve_once |

Workflow-run tools expose Graph Engineering and product-loop runs to AI agents with a hard authority boundary: the host sets source: "ai" on every submitted signal, and only AI-artifact event types are accepted. Human events (approvals, rejections, retries, cancellations) never pass through MCP — they belong to the swarm-dao CLI human channel.

Programmatic API

import { createSwarmDaoMcpServer, ensureDaoStorage, createMcpHostAdapter } from "@guyghost/swarm-dao-mcp";

await ensureDaoStorage("/path/to/repo");
const server = createSwarmDaoMcpServer("/path/to/repo");
await server.connect(yourTransport);

How deliberation works over MCP

MCP hosts cannot have the server spawn sub-agents, so Swarm DAO uses manual deliberation here:

1. dao_deliberate returns a dispatch plan (agents + models).
2. The host spawns one sub-agent per plan entry itself.
3. dao_record_outputs feeds the sub-agent outputs back into the DAO.

License

MIT

Long-running tools

dao_improve_once executes real worker agents (herdr) during the sampling and auditing phases — a single call routinely takes one to several minutes. MCP clients default to a 60-second request timeout, which kills the call mid-flight (leaving the series state untouched, but orphaning the spawned herdr workspace; the next call cleans it up). Hosts driving improvement series should raise their request timeout:

// per call
const result = await client.callTool(
  { name: "dao_improve_once", arguments: { seriesId: "…" } },
  undefined,
  { timeout: 15 * 60_000 },
);
// or globally
const client = new Client(info, { defaultRequestTimeout: 15 * 60_000 });