Your AI Agents, In Your Pocket
A native iOS AI Agent app — multiple customizable AI assistants, right on your phone

Website · Features · Quick Start · MIT License · 中文文档

Most AI chat apps are just chat boxes. iClaw is different — it's a full-blown AI Agent platform where every agent has its own soul, memory, and skills, and can execute code directly on your device.

iClaw is inspired by OpenClaw — the open-source personal AI assistant. We share the same vision of giving users a truly personal, customizable AI agent. Some shared DNA:

• Soul-driven agents — OpenClaw pioneered the Brain-Body-Soul model; iClaw adopts a similar philosophy with SOUL.md / MEMORY.md / USER.md mind files
• Multi-provider LLM — both let you swap models freely without losing context
• Cron scheduling — both support scheduled agent tasks
• Function calling & tool use — agents can take real actions, not just chat

In short: OpenClaw is your AI command center on a server; iClaw puts the same power directly into your pocket — no infrastructure required.

• Create multiple AI agents, each with its own SOUL.md (personality), MEMORY.md (persistent knowledge), and USER.md (user profile)
• Agents can spawn sub-agents that inherit the parent's soul configuration
• Flexible Markdown config space — you define how each agent behaves

• JavaScript interpreter — executed in a WKWebView sandbox with automatic crash recovery
• Two execution modes: repl (expression evaluation) and script (full scripts)
• Save and reuse code snippets — build your mobile code library

• Session-centric interaction design
• SwiftData persistence — your data is safe and never lost
• Auto context compression — when token limits are exceeded, older messages are compressed and summaries are injected into system prompts

Agents don't just chat — they remember. iClaw includes a built-in Retrieval-Augmented Generation (RAG) framework that turns past sessions into a searchable knowledge base, so agents can recall and reason about prior conversations.

• On-device vector embeddings — powered by Apple's NaturalLanguage framework (sentence-level with word-vector fallback), all computed locally with zero cloud dependency
• Hybrid search — vector similarity search (cosine, threshold >0.3) with automatic keyword fallback, ensuring results are always available even without embeddings
• Agent-callable tools — two function-calling tools exposed to agents:
  • search_sessions — semantic search across an agent's session history
  • recall_session — retrieve and paginate through a specific session's messages (token-aware, max 4 000 tokens)
• Automatic context injection — related sessions are discovered at session start and injected into the system prompt, so agents have relevant history without being asked
• Background indexing — embeddings are lazily backfilled on app launch in batched, CPU-throttled passes so the UI stays smooth
• Privacy & ownership — embeddings stay on-device; each agent can only search its own sessions; the current session is auto-excluded to prevent self-reference

• Compatible with any OpenAI API-compatible endpoint
• Full SSE streaming support with real-time typewriter output
• Built-in presets: OpenAI / DeepSeek / OpenRouter / Ollama (local)
• Configure multiple providers and switch freely

• Built-in browser powered by WKWebView — browse the web without leaving the app
• Agent-driven automation: agents can navigate pages, click elements, fill forms, extract data, and execute JavaScript — all through function calling
• 9 browser tools: browser_navigate, browser_click, browser_input, browser_extract, browser_execute_js, browser_get_page_info, browser_select, browser_wait, browser_scroll
• Mutex lock ensures only one agent session controls the browser at a time; users see a live banner and can take over at any time
• Simplified DOM extraction gives agents a readable view of page structure (links, forms, buttons, text)

• read_config / write_config — read and modify agent mind files
• execute_javascript — run code
• save_code / load_code / list_code — manage code snippets
• create_sub_agent / message_sub_agent — sub-agent lifecycle management
• schedule_cron / unschedule_cron — scheduled task management
• search_sessions / recall_session — session RAG & cross-session retrieval
• browser_* — 9 browser automation tools (see In-App Browser section above)

• Installable skill library to give agents new capabilities
• Flexibly attach skills — one skill can serve multiple agents

• Create scheduled tasks for agents with automatic background execution
• Built on BGTaskScheduler for system-level reliable scheduling
• Deep Link triggers: iclaw://cron/trigger/{jobId}

• Xcode 15+
• iOS 17.0+
• XcodeGen

# Install XcodeGen
brew install xcodegen

# Generate the Xcode project
cd iClaw
xcodegen generate

# Open in Xcode
open iClaw.xcodeproj

1. Open the app → Settings → Add an LLM provider (enter API endpoint, key, and model name)
2. Go to Agents → Create your first agent
3. Go to Sessions → Start a new session → Chat away!

iClaw/
├── project.yml              # XcodeGen config
├── iClaw/
│   ├── App/                 # App entry point & root navigation
│   ├── Models/              # SwiftData models
│   ├── Services/
│   │   ├── LLM/            # LLM API client, SSE parser
│   │   ├── Agent/           # Agent management
│   │   ├── Session/         # Session service, context compression
│   │   ├── Prompt/          # System prompt construction
│   │   ├── FunctionCall/    # Tool definitions & routing
│   │   ├── CodeExecution/   # Code execution engines
│   │   ├── Browser/         # In-App Browser service & automation
│   │   ├── CronJob/         # Cron job scheduling
│   │   └── Skill/           # Skill management
│   ├── ViewModels/          # Observable view models
│   ├── Views/               # SwiftUI views
│   └── Resources/           # Assets & default templates

This project is open-sourced under the MIT License.

Copyright (c) 2026 ShadowMov

iClaw — Your AI Agents, In Your Pocket.
iclaw.shadow.mov