Tell it what to do — and it actually gets it done.
Draft docs, write code, install software — whatever you need.
Windows · macOS · Linux
Four core modules covering every AI collaboration scenario.
18+ LLM providers (Anthropic, OpenAI, DeepSeek, etc.) and 16+ platforms (WeChat, DingTalk, Feishu, Telegram, Discord, etc.). Multi-turn conversations, autonomous decisions, 4-level security engine, MCP tool integration.
Provider names indicate compatible connections only. ClawSky is independent and is not affiliated with, endorsed by, or sponsored by these model providers.
Local document knowledge base + MCP Server. PDF/Word/Excel/PPT parsing, ONNX vector embeddings, DuckDB storage, OCR recognition.
Cherry Markdown editor. Sub-second startup, .md file association. Imports DOCX, XLSX, XLS, PPTX. Exports PDF and DOCX.
Infinitely extensible via Model Context Protocol. Knowledge base runs as an isolated MCP Server. Browser automation, desktop control, API calls — everything via MCP.
Not a marketing buzzword — it's knowledge base + audit logs + feedback loops, already built in.
Agent handles tasks, calls MCP tools, makes autonomous decisions in multi-turn conversations.
Full-chain audit logs capture every decision path, tool call, and result.
Extract effective patterns from operation history and structure them into the local knowledge base.
Agent autonomously retrieves relevant experience via semantic search and injects it into the next decision.
All experience data stored locally. Your operation history never leaves your device.
Every decision is logged. Review, edit, or delete any stored experience anytime.
The AI decides when and what to retrieve — no manual configuration needed.
Same semantic search engine — vector embeddings + jieba tokenization, not keyword matching.
Not a danger-command blacklist bolted on as an afterthought. Every tool call passes through one security engine first — and you pick how much freedom the agent gets, from read-only to unrestricted.
Only low-risk, read-only tools run — file reads, directory listing, search. Nothing can write, execute, or reach out.
Read-only opsFile edits inside your workspace plus whitelisted shell commands. Out-of-bounds writes and unknown commands are blocked automatically.
Workspace + whitelistThe agent can do everything — but high-risk moves wait for your one-tap approval: deletes, installs, network calls, system changes, process control, mouse & keyboard.
High-risk needs confirmationEverything runs, no prompts. For trusted environments and power users who want maximum autonomy.
No promptsEvery tool call is classified, gated, confirmed, and logged at a single chokepoint — not a patchwork of ad-hoc filters.
Risky operations pause for an in-app or terminal prompt. Approve once, or allow that category for the whole session.
Every decision, tool call, and risk category is recorded — so you can always review what the agent actually did.
Security policy and audit logs live on your device. Workspace boundaries and an isolated shell environment keep secrets in.
Start free, grow as you need. Some components planned open source.
Coming soon. Core components are planned as MIT open source.
AI Agent framework, forked from nanobot (MIT). Multi-turn chat, tool calls, skills, multi-channel, MCP integration, security engine.
Lightweight Markdown desktop editor. Cherry Markdown + PyWebView. Sub-second startup, .md file association. Imports Office files (DOCX, XLSX, XLS, PPTX) into Markdown, exports to PDF or DOCX.
Bidirectional document ↔ Markdown converter. Imports DOCX, XLSX, XLS, PPTX; exports DOCX and PDF. pip install mdport