Local-first agent framework I've been building solo. Default backend is Ollama (qwen3 models), with vLLM/OpenAI/Anthropic/Gemini + 14 others as opt-in. Runs on Windows, Linux, Mac, plus Flutter Command Center for Android/iOS/Web.
What's in it:
- 127+ MCP tools across 30+ modules (memory, kanban, web research, file ops, code execution)
- 16 channels — CLI, Telegram, Discord, Slack, WhatsApp, Voice, WebUI, more
- PGE-Trinity orchestration (Planner / Gatekeeper / Executor) with risk classification on every tool call
- Pack-plugin system with EULA click-through + SHA-256 hash pinning
- 6-tier memory + audit-chain JSONL + Trace-UI (live WebSocket view of the agent loop)
The unusual part:
- ~14 500 tests, 89% coverage gate, mypy --strict, spec-first development
- Heuristic constants live in YAML, not source
- Currently shipping a neuro-symbolic program synthesis engine over ARC-AGI-3: Phase 1 done with measured uplift, Phase 2 (LLM-prior + MCTS + 3-zone refiner) landing this sprint
pip install cognithor — Apache 2.0, code at github.com/Alex8791-cyber/cognithor
Curious what you'd want it to do that it can't yet.