You know the pattern: upload a CSV to ChatGPT, ask for “analyze this,” and you get a polite paragraph that says nothing. No charts. No actionable output. Just a summary of what it would do if it had access. LobsterAI — 5,400 stars on GitHub, the first open-source desktop AI agent from NetEase AI’s Youdao division — doesn’t have that problem. It connects to your real desktop: files, terminal, browser, local projects. And it actually executes.

What Makes Cowork Different

The headline feature here is Cowork mode. Instead of generating text about what it would hypothetically do, LobsterAI opens a bridge to your actual working environment. Give it a spreadsheet, and it’ll write a Python analysis script, run it locally, and generate a visualization page. Give it a folder of PDFs, and it’ll batch-process them into a structured report. And every file-accessing tool call gates behind your approval — I got a permission dialog asking “LobsterAI wants to read /Users/me/data/sales.csv” before anything touched my disk.

Under the hood, it runs on OpenClaw — their custom open-source agent framework that ships with 28+ built-in skills. Web search, docx/xlsx/pptx generation, video creation via Remotion, browser automation via Playwright, image generation via Seedream, stock analysis — the list is long. Plus it supports the MCP protocol, so you can plug in external tools the same way you would with Claude Desktop. So the OpenClaw agent framework is what makes this different from every other “AI that browses your files” demo. If you’re into agent toolkits, I covered a similar ecosystem in the Composio review.

The Phone-Command-Your-PC Trick

This is the part that made me stop scrolling the README and actually install it. LobsterAI bridges to 7 IM platforms: WeChat, WeCom, DingTalk, Feishu, QQ, Telegram, and Discord. You send a message from your phone, the desktop agent picks it up, executes, and fires the result back.

So I set up the Telegram bridge. Sent it: “Research the global AI Agent market and turn traffic-report.pdf into a PPT deck.” And it came back six minutes later with a generated slide deck on my desktop and a summary pushed to my phone. Did it while I was making coffee. That’s the kind of “AI assistant” I signed up for.

What I Actually Ran

Install flow is straightforward if you’re comfortable with npm — clone, npm install, then npm run electron:dev:openclaw for the first launch. But the first build takes a while because it clones and compiles the OpenClaw runtime. On my Ryzen 9 workstation it took about three and a half minutes. After that, subsequent launches via npm run electron:dev are instant — it reuses the cached runtime.

My first task was simple: “Analyze the product-growth.xlsx in my Downloads folder and build me a visualization page.” Honestly, I didn’t expect it to find the right file on the first try. But it did — correctly, not a cached path, not a sandbox — wrote a Python script with matplotlib, generated an HTML dashboard, and opened it in my browser. That’s about 12 seconds of actual execution time after the initial agent warmup.

What To Watch For

Still, it’s not a finished product. 835 open issues on the repo tells you the team is building and shipping fast — but you’ll hit rough edges. The Electron app is RAM-hungry (about 400 MB idle). And the Cowork permission dialog appears every single time for file access unless you approve the session — which is good for security but can get annoying during long workflows. Plus some of the IM integrations (WeChat, QQ) require Chinese-platform accounts that your average Western developer won’t have. Telegram and Discord work fine though.

Bottom Line

Yet LobsterAI is the first genuinely desktop-grade open-source AI agent from a major vendor. If you’re tired of chat interfaces that promise desktop access and deliver nothing, this is the real thing. Sure, it’s rough, it’s RAM-heavy, and the first build takes a minute — but the Cowork architecture is right. Your desktop, your tools, your data, actually doing work. For another take on AI agent infrastructure from a Chinese tech giant, check out my OpenViking review.