tombarr/iclaw
iClaw - local, private, macOS AI assistant
What's novel
iClaw runs completely free and fully local on your mac without internet or API keys
Code Analysis
15 files read · 4 roundsA local, on-device AI agent for macOS/iOS that routes user queries through a multi-stage pipeline (NLP preprocessing, ML classification, tool routing, multi-step agent execution) using Apple Foundation Models or Ollama, with progressive memory, guardrails, and 6 specialized ML classifiers.
Strengths
Exceptionally deep implementation with a well-designed FSM-based execution engine, multi-stage tool routing with ML classifiers, three-tier progressive memory, and sophisticated NLP preprocessing (spell correction with brand-name protection, multilingual entity extraction, language-neutral conversational gating). The code is heavily documented with design rationale, uses Swift 6 strict concurrency throughout, and has extensive test injection points.
Weaknesses
The ExecutionEngine at 138KB is a monolithic actor that, while well-organized, concentrates enormous complexity in a single type. The project targets macOS 26+/iOS 26+ (not yet released), limiting practical usability. Some of the ML classifier training pipeline lives in a separate MLTraining directory with Python/Swift scripts that feel somewhat disconnected from the main codebase.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
29
Contributors
3
Files
36
Active weeks
1
Repository
Language
Swift
Stars
1
Forks
0
License
MIT