Free, local AI Agent for macOS and iOS 26+
iClaw runs on-device using Apple Intelligence or Ollama

Website · Features · Tools

iClaw is a local, native AI agent for macOS 26+ (and soon iOS 26+) written in Swift. It runs on-device using Apple Intelligence or Ollama. No usage limits, API keys, or subscriptions.

Why iClaw? iClaw envisions a helpful AI experience where you get the information you need using the hardware you already have, all without sacrificing your security or privacy.

Podcast playback and research	Time, weather, and history

iClaw lives in the menu bar as a floating Liquid Glass HUD, or optionally as a dock application.

• Private. All AI runs completely on-device, not in the cloud
• Free. No API keys, usage meters, or monthly subscriptions
• Easy. It's just an app, no custom models to configure
• Green. Use a fraction of the energy that cloud agents require
• Open Source. Free to review, contribute, or fork

Dynamic widget generation	Conversational follow-up	Settings and browser bridge

• 100% On-Device — All inference runs locally using Apple Intelligence or Ollama. Network is only used for fetching data (weather, web search, stock quotes).
• Secure Design.- iClaw lives in the App Sandbox and only has access to the tools, data, and accounts that you share with it.
• Explicit Consent — Destructive or non-reversible actions always require explicit user confirmation.
• Extensible Skills — Add your own skills to teach iClaw new capabilities.
• 4K Token Budget — Every prompt component has an explicit token budget. The system is tuned to not exceed the on-device model's context window.
• Progressive Memory — Three-tier memory (working facts, running summary, vector archive) compresses, structures, and anchors recall for natural follow-ups.
• Adaptive Routing — ML classifier gates complexity: high confidence → fast single-tool path, medium → @Generable plan decomposition, low → multi-turn agent reasoning loop.
• Multi-Step Agent — Complex queries are decomposed into typed execution plans that the agent runner executes with inter-step reasoning.
• ReAct Validation Loop — Ingredients are validated for relevance. If a misroute is detected, the system re-routes with tool suppression.
• LaTeX Rendering — Mathematical expressions in LLM responses are detected and rendered natively via SwiftMath.
• Soul-driven — iClaw uses a (very concise) SOUL.md to generate a USER.md from interaction history.

• Mac App Store: publish iClaw on the Apple App Store for macOS
• iOS App: run iClaw natively on any iPhone with Apple Intelligence, published on the Apple App Store for iOS
• Continuity: take iClaw on the go and automatically sync messages, files, and actions between Mac and iPhone
• Learned Skills: learn what actions you take repeatedly and turn these into repeatable, personalized Skills
• Web Extensions: access web pages and take actions on Chrome, Firefox, or Opera through web extensions similar to Safari
• Email: expand email support to any inbox, not just Mail.app. Support MailKit for automatic triage and categorization.
• Dynamic Widget Catalog: generate a catalog of well-designed slot-fill widgets to go beyond the wall of text

The MAS build uses iClaw-MAS.entitlements which excludes Apple Events and CloudKit. Tools that depend on AppleScript (i.e. AutomateTool, ReadEmailTool) are compile-time excluded via #if !MAS_BUILD. Other tools (SystemControl, Shortcuts) have MAS-compatible fallbacks using Core Audio, URL schemes, and NSWorkspace. Productivity tools (Calendar, Messages, Reminders, Notes, Contacts, Email) use a soft-creation fallback pattern: on permission denial or error, they show a preview widget with action buttons (e.g., "Add to Calendar" via .ics file, "Open in Messages" via sms: URL, "Add to Contacts" via .vcf vCard, "Copy & Open Notes" via clipboard).

Built-in: Books, Countries, Crypto, Emoji, Horoscope, Movies, Quote, Research, Rubber Duck, Tech Support.

Three execution paths based on query complexity:

Input → Preprocessing (NER + Spellcheck) → ML Classifier
  ├── High confidence (≥0.90) → FAST PATH: Single tool → Widget → Output
  ├── Medium (0.35–0.90) → PLANNING: @Generable plan → Sequential execution → Output
  └── Low (<0.35) → AGENT: Multi-turn reasoning loop with domain-scoped tools → Output

The fast path preserves the existing low-latency pipeline for simple queries ("what's the weather"). The planning path uses @Generable structured output for type-safe tool decomposition. The agent path enables multi-step reasoning for complex queries ("book a walking lunch with Sarah the next sunny day").

1. Attachment hint (file type detection)
2. Follow-up detection (3-layer: slots, ML, NLP)
3. Tool chips (#weather, #stocks, etc.)
4. Ticker symbols ($AAPL)
5. URL detection
6. Skill matching
7. Synonym expansion
8. ML classification (MaxEnt CoreML)
9. Heuristic overrides (entity-aware suppression for Calculator/Convert/Random/SystemInfo/Stocks)
10. LLM fallback
11. Conversational

Tools return raw data as ingredients. When no tool widget is returned, a separate LLM call generates a visual layout using a pipe-delimited DSL:

H|title|subtitle            — Header
S|value|label               — Hero stat
SR|val1;label1|val2;label2  — Stat row
KV|key|value                — Key-value pair
L|title|subtitle|trailing   — List item
TB|col1|col2|col3           — Table header
TR|val1|val2|val3           — Table row

Widgets are validated post-generation: placeholder values removed, single-block widgets rejected, table units hoisted to headers, minimum content thresholds enforced.

Three-layer guard prevents overcorrecting brand names and proper nouns:

1. Capitalization — Mid-sentence capitalized words are skipped
2. NER adjacency — Words near NER-identified entities are protected
3. Common word filter — Corrections accepted only toward top-5000 English words

Three-tier progressive memory within the 4K token budget:

• Tier 1 — Working Facts (5 slots, ~50 tokens): Structured facts from tool results, scored by recency + entity overlap
• Tier 2 — Running Summary (~80 tokens): Incremental LLM fold of evicted facts
• Tier 3 — Vector Archive: NLEmbedding cosine similarity for long-term semantic retrieval

iClaw previously shipped a LoRA adapter on top of Apple's on-device Foundation Models. The adapter has been removed and is not wired up pending resolution of this Apple Developer Forums thread. Until that issue is resolved, iClaw runs against the base SystemLanguageModel without an adapter.

Requires macOS 26+ (or iOS 26+) with Apple Intelligence enabled.

make build     # Debug build
make run       # Build + open
make test      # Run all tests