使用说明
swarm-mcp
Durable local coordination between independent coding-agent sessions: acknowledged messages, fenced task ownership, shared context and retained evidence. Native runtimes own their children; Swarm coordinates across session boundaries.
Get started
The candidate uses one Node owner per isolated database/profile and nine MCP tools. Trusted launchers enroll sessions and supply capabilities; MCP adapters do not open the database or select their own identity.
- Install and enroll
- Swarm API and consumer skill
- Host support and limitations
- Embed in a runtime or dispatch through Herdr
- Storage limits and offline maintenance
This is a local release candidate. Building or installing it does not migrate an existing swarm or switch live host configuration.
Boundaries
Swarm owns durable acceptance, acknowledgment, attempts, reservations and evidence. Host adapters own safe context delivery; workspace providers own execution. Trackers retain human-facing work. A successful send does not prove model processing, and a lease cannot guarantee exactly-once external side effects.
Automatic delivery varies by host. File reservations are cooperative. One local OS user/profile is the trust boundary; this is not a distributed or hostile multi-tenant service. See the architecture decision.
Entrypoints
swarm-mcp and swarm-coordinator-mcp serve the same Swarm API. Both require
the endpoint and session capability supplied by a trusted runtime launcher.
See runtime embedding for enrollment and launch options.
Each profile uses its own coordinator database. See the application boundary decision.
Development
bun install --frozen-lockfile
bun run check
The focused coordination gate also measures actual 32-agent MCP context:
python3 -m pip install tiktoken==0.12.0
PYTHON=python3 bun scripts/verify-coordination.ts
npm run verify:package
See verification, benchmark evidence, and packaging.
