The open-source desktop AI agent that does real work on your computer.
Describe a task in plain language. LobsterAI reads your files, runs tools, drives a browser,
edits Office documents, and hands back finished work — at your desk or from your phone.
English · 简体中文
Most AI assistants stop at an answer. LobsterAI is built to finish the job.
- It works where your work lives. Point it at a folder and it reads spreadsheets, writes documents, runs scripts, and browses the web, asking before anything sensitive.
- You see the result, not just a reply. Web pages, charts, images, videos, PDFs, and Word, Excel, and PowerPoint files open in a side panel, ready to edit or share.
- It keeps working when you step away. Schedule recurring jobs, or message your desktop agent from WeChat, Feishu, DingTalk, Telegram, and more.
- It's yours. MIT-licensed, with sessions stored locally. Sign in to use the built-in models, or bring your own API keys.
Built by NetEase Youdao, LobsterAI is the first open-source desktop-grade agent from a major Chinese tech company.
Native Office editing |
Kits, skills, and connectors |
Scheduled tasks |
Message it from anywhere |
| Capability | What you get |
|---|---|
| Cowork sessions | Long-running tasks on local projects with streaming progress, a tool timeline, session forks, and approval prompts before file, shell, or network actions. |
| Artifacts and sharing | Live previews for HTML, SVG, Mermaid, images, video, Markdown, code, PDF, and Office files. Edit images in place and share results as a link. |
| Built-in browser | An in-app browser the agent can drive, with tabs, saved logins, and passkey support, for research and web workflows. |
| Computer use | Let the agent operate desktop apps on Windows x64 and macOS 13 or later. |
| Multiple agents | Create agents with their own identity, model, skills, and working folder, add ready-made experts from the Expert Plaza, and bind agents to chat channels. |
| Skills | Built-in skills for documents, spreadsheets, slides, PDFs, web search, browser automation, image and video generation, stock research, writing, email, and more. See SKILLs/. |
| MCP connectors | Add local or remote MCP servers from the marketplace, sign in with OAuth, and choose which tools each server exposes. |
| IM channels | WeChat, WeCom, DingTalk, Feishu/Lark, QQ, NetEase IM, NetEase Bee, POPO, Telegram, Discord, and email. Multi-instance platforms can route each bot to a different agent. |
| Desktop companion | A floating orb for quick tasks, screenshots, and actions on selected text, such as translating or reading aloud. |
| Memory and Dreaming | Durable memory files, plus optional scheduled "Dreaming" runs that consolidate what the agent learned, so preferences carry across sessions. |
| Models | Built-in models (DeepSeek, GLM, Kimi, Qwen, MiniMax, and more) after sign-in, or your own keys for OpenAI, Anthropic, Gemini, xAI, OpenRouter, Ollama, LM Studio, and any OpenAI-compatible endpoint. |
| Library | Every file your agents create, collected in one place across sessions, both local and shared. |
| Scenario | Example prompt |
|---|---|
| Build a local tool | "I still track inventory and sales in Excel. Build a local inventory system that records purchases and sales, calculates stock and profit, and opens in my browser." |
| Analyze local data | "Use product-growth.xlsx to build a visual dashboard and summarize the main growth drivers." |
| Make a deck | "Research the AI agent market and turn the findings into a presentation." |
| Check a web dashboard | "Open the ads dashboard every morning, check spend and conversion anomalies, and summarize likely causes." |
| Screen documents | "Turn the resumes in this folder into a screening sheet and shortlist the strongest candidates against the JD." |
| Run scheduled work | "Every weekday at 9 AM, collect yesterday's AI news and send me a concise digest." |
- Install LobsterAI. Download it from the official website, or get the latest release directly: macOS (Apple Silicon) · macOS (Intel) · Windows x64
- Choose a model. Sign in to use the built-in models, or add your own API key under Settings → Custom Model.
- Give it a task. Pick a working folder, describe what you want, and approve actions as it works. Generated files open in the side panel.
- Go further. Install kits from Kits · Skills · Connectors, set up recurring work in Scheduled Tasks, or connect a chat app under Settings → IM Bot.
flowchart LR
subgraph App["LobsterAI desktop app (Electron)"]
UI["Renderer<br/>React · Redux · Tailwind<br/>sessions · artifacts · Office editors"]
Main["Main process<br/>Cowork sessions · permissions · SQLite<br/>skills · MCP · IM · scheduler"]
UI <-->|"preload IPC"| Main
end
Main <-->|"config sync · events"| Gateway["OpenClaw gateway<br/>bundled agent runtime"]
Gateway --> Models["LLM providers"]
Gateway --> Tools["Local tools<br/>files · shell · browser<br/>computer use · MCP servers"]
Chat["Chat apps<br/>WeChat · Feishu · DingTalk · Telegram …"] <--> Gateway
- Cowork is LobsterAI's product layer. It owns sessions, messages, permissions, artifacts, agents, and IM bindings, and stores them locally in SQLite.
- OpenClaw is the agent runtime underneath. LobsterAI ships a pinned build, renders its config from app state, and translates its events into the Cowork stream.
- The renderer is sandboxed with context isolation and talks to the main process only through the preload IPC bridge.
Requirements: Node.js >=24.15.0 <25, npm >=11.17.0 <12, git, and pnpm.
git clone https://github.com/netease-youdao/LobsterAI.git
cd LobsterAI
npm install
# First run: fetches the pinned OpenClaw version into ../openclaw, builds the runtime, and starts the app
npm run electron:dev:openclaw
# Later runs, once the runtime exists
npm run electron:devCONTRIBUTING.md covers tests, linting, runtime options, packaging installers, and the project layout.
- Sessions, messages, and settings are stored locally in
lobsterai.sqliteunder the app's user data directory. Agent memory and runtime state live alongside it underopenclaw/. - File operations, shell commands, and network access are permission-gated, and the agent asks before it acts.
- Renderer windows run with context isolation, sandboxing, and no Node.js integration.
Join the WeChat group for help, feedback, and release news. You can also follow @LobsterAIYoudao on X.
Report bugs and request features through GitHub Issues.
Contributions are welcome. Read CONTRIBUTING.md to set up your environment, and look for good first issues to get started.
MIT © NetEase Youdao