Briefly
A local agent dashboard for the boring but important parts: task intake, sandboxes, review, and model health.
Local Swarm is a Gemma-only local agent system for development work on an Apple Silicon workstation. It combines an orchestrator API, dashboard, CLI, SQLite run history, model routing, tool policy, review queues, and sandbox-aware automation.
The important design choice is that autonomous work happens inside disposable worktrees or copied sandboxes. Agents prepare a reviewable packet; merge back requires explicit approval and gate checks.
This is the part of local AI work I care about most: not the chat window, but the controls that make agent work inspectable.
What I built
- Dashboard covers chat, task intake, agent runs, review queue, model health, tool policy, memory, benchmarks, and review files.
- Sandbox policy keeps main-worktree writes blocked until an approved merge operation.
- Review queue and merge gate make agent output inspectable before it affects source.
- Model health keeps the local LM Studio/Gemma runtime visible instead of hidden behind a generic chat box.
A few notes
- The dashboard covers task intake, review queue, an agent run sandbox, and model health.
- The README documents Fastify orchestration, React/Vite dashboard, CLI commands, SQLite history, approval gates, and tool allowlists.
- The system is explicitly local-first and Gemma-family only for LLM routing.