A local-first micro-agents runtime. One resident planner LLM receives intents and spawns short-lived, sandboxed micro-agents to do the work. Each agent gets exactly two tools, search and execute; everything else it needs lives on the filesystem and is reached through them. Sandboxing is a ladder (process, docker, microVM) so you pay isolation cost only when the threat model demands it. Every observable event is one line in a per-run JSONL audit log: if cat and grep cannot see it, it does not exist as state. Memory is plain files, no database, no vendor SDK, no cloud lock-in. Ollama by default, any OpenAI-compatible backend as fallback.
- GIS: CLI agent that resolves GitHub issues end to end, RAG knowledge base over the repo, patch generation, PR creation
- browserstack-agent-edge-cases: three reproducible production failure modes for browser agents