SafeClaw runs entirely in your browser tab — multiple LLM providers, built-in tools, encrypted storage, and offline-capable local models. No server, no tracking, full privacy.
Choose the best model for your needs. Cloud for quality, local for privacy and offline use. Quality-first routing picks the right provider automatically.
Claude Opus 4.6, Sonnet 4.6, and Haiku 4.5 — industry-leading reasoning with extended thinking and tool use.
CloudGemini 2.5 Pro and 2.0 Flash via the Gemini API — fast, capable, with native function calling.
CloudRun Qwen3 models locally via WebGPU. No data leaves your device. Works fully offline.
Local — WebGPUChrome's built-in AI model. Zero download, instant availability for summarization tasks.
Local — Built-inEverything runs in a single browser tab. Your data never touches a server you don't control.
No backend, no telemetry, no tracking. API keys are encrypted with AES-256-GCM using non-extractable Web Crypto keys stored in IndexedDB.
Install as a Progressive Web App on any device. Works on desktop, tablet, and mobile. Feels native.
The AI can read/write files, run code, fetch URLs, manage tasks, and update its own memory — autonomously chaining tools to complete complex requests.
With WebLLM local models, SafeClaw works without an internet connection. Quality-first routing automatically falls back to local when offline.
Quality-first provider selection: cloud by default for best results, automatic fallback to local models when offline or rate-limited.
Conversations, files, and tasks persist in IndexedDB and OPFS. Your assistant remembers context across sessions.
Connect a Telegram bot to chat with your assistant from any device. Pure HTTPS — no WebSockets or special protocols required.
Schedule recurring tasks with cron expressions. Your assistant runs them automatically in the background while your tab is open.
Built-in file manager backed by Origin Private File System (OPFS). Read, write, and browse files in per-group workspaces.
Personalized AI use-case suggestions based on your profile. Discover 20+ curated workflows across automation, research, and development.
Discover what you can build with SafeClaw. Top use cases from the community, sourced from awesome-openclaw-usecases.
Aggregate and summarize tech news from multiple sources into a curated daily briefing delivered on your schedule.
Summarize newsletters and send a condensed digest via email. Cut through the noise automatically.
Automatically discover and track contacts from email and calendar. Never lose touch with important connections.
Convert meeting transcripts into structured summaries with automatic task creation and follow-ups.
Automate idea scouting, research, and content tracking for video or blog production workflows.
Build a searchable RAG-powered knowledge base from URLs, tweets, and articles you collect.
Scan GitHub, HN, npm, PyPI, and Product Hunt before building — know if your idea already exists.
Track food intake and symptoms to identify triggers, with scheduled reminders and trend analysis.
Daily check-ins tracking habits with adaptive tone, accountability, and streak monitoring.
Eight tools the AI can chain together autonomously to accomplish complex tasks.
No magic, no hidden complexity. A transparent pipeline from message to response.
From message to response in six steps — all within your browser.
Your message is captured by the React UI and dispatched to the orchestrator.
The router selects the best provider based on availability, your preference, and network status.
The orchestrator builds the prompt with system instructions, conversation history, memory, and tool definitions.
The payload is sent to a dedicated Web Worker that runs the LLM call off the main thread.
If the model requests tool use, the worker executes tools and re-invokes the model — up to 25 iterations.
The final response is posted back to the main thread, stored in IndexedDB, and rendered in the chat.
SafeClaw defaults to cloud models for best quality, and gracefully falls back when needed.
| Condition | Provider Selected |
|---|---|
| User explicitly selects a provider | User's choice (any provider) |
| Local preference set to "always" | WebLLM or Chrome AI |
| Device is offline | WebLLM (if downloaded) |
| Cloud API returns 429 (rate limit) | Alternate cloud → local fallback |
| Default (online, no issues) | Configured cloud provider (Anthropic or Gemini) |
SafeClaw is open source under the MIT License. Inspect every line, fork it, make it yours.