rah757/ProductivityClaw
Local-first AI productivity agent — calendar, memory, proactive suggestions. Powered by LangGraph + Qwen 3.5 35B A3B.
What's novel
Local-first AI productivity agent — calendar, memory, proactive suggestions. Powered by LangGraph + Qwen 3.5 35B A3B.
Code Analysis
18 files read · 5 roundsA local-first AI agent that integrates with Apple's ecosystem (Calendar, Mail, Notes) via EventKit/ScriptingBridge, uses a local MLX LLM (Qwen 3.5) through LangGraph for stateful tool-calling conversations, extracts persistent facts from conversations, and proactively sends briefings via Telegram wi
Strengths
Genuinely well-architected local-first agent with a real memory system (fact extraction pipeline with confidence scoring and staging), a thoughtful proactive heartbeat that respects MLX's single-threaded nature via a user-activity lock, and production-quality Telegram streaming with tiered flush intervals and sentence-boundary snapping. The skill plugin system (manifest.json + execute.py) and the human-in-the-loop write confirmation flow are clean, practical patterns.
Weaknesses
The legacy Ollama-based agent.py is dead code that should be removed, and the streaming implementation in graph_agent.py has a subtle bug where it consumes the first streaming pass to detect tool calls then re-runs the entire graph non-streaming, wasting an LLM call. The intent_router is a no-op that adds indirection without value, and the _inject_trace_id function uses exec_module to mutate module globals, which is fragile.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
25
Contributors
1
Files
70
Active weeks
4
Repository
Language
Python
Stars
1
Forks
1
License
—