LobsterAI: NetEase Youdao's Desktop Agent Built on the OpenClaw Runtime

The problem LobsterAI targets: agents that only talk versus agents that operate

Most chat agents stop at text. LobsterAI is positioned as a desktop agent that works inside the local environment: files, terminal commands, browser workflows, documents, spreadsheets, slides, IM channels, scheduled jobs and project workspaces. The README's own framing is an "All-scenario office assistant Agent" that runs tools on the real desktop. The audience is therefore narrow and specific: people whose work lives in local files and office formats, not teams looking for a server-side automation framework. The repository lists macOS and Windows only, and the README's example prompts are all local-work prompts, such as building a browser-opened inventory system from Excel data, turning a folder of resumes into a screening sheet, or collecting AI news every weekday at 9 AM. That is a coherent scope. It also means anyone expecting a Linux daemon or a REST API should stop reading here.

Cowork on top, OpenClaw underneath: the actual split

The architecture is a two-layer split, and understanding it explains most of the project's constraints. Cowork is the product and session layer; OpenClaw is the runtime and gateway. The README states that this split keeps local persistence, permissions, UI state, artifacts, agents, memory and IM bindings in the desktop app while OpenClaw handles agent execution. The repository layout matches that description. The renderer is React 18 with Redux Toolkit and Tailwind, holding artifact renderers, settings, agent and session UI, skills, MCP configuration, scheduled tasks and IM configuration. The Electron main process owns lifecycle, IPC, SQLite persistence, auth, logging, OpenClaw startup, runtime repair, skill sync, IM gateways and artifact services. Four named modules bridge the two: openclawEngineManager, openclawConfigSync, openclawRuntimeAdapter and coworkEngineRouter, which the README says translate LobsterAI state into OpenClaw runtime behavior. Persistence is deliberately local. Sessions and app data live in SQLite, and workspace memory uses files such as MEMORY.md, USER.md, SOUL.md and daily notes. The trade-off is visible: you get durable local context and no server dependency, but nothing in this description suggests a multi-user or hosted deployment model.

Installing LobsterAI and running a first scheduled task

For normal use, the README points to installers rather than a package manager. Download the latest macOS or Windows installer from the official website or from GitHub Releases. No npm or Homebrew install path is documented, so do not expect one.

Running from source has hard prerequisites. Node.js must satisfy >=24.15.0 <25, and the first run needs git and pnpm to build the pinned OpenClaw runtime from a sibling ../openclaw checkout. The clone and install step is ordinary.

git clone https://github.com/netease-youdao/LobsterAI.git
cd LobsterAI
npm install

The first development run is not the same command as daily development. This one builds the pinned OpenClaw runtime and should be used once.

npm run electron:dev:openclaw

After that runtime exists, the shorter command is the daily path. The renderer dev server listens on http://localhost:5175, which is the address to open in a browser if the Electron window does not appear.

npm run electron:dev

For a first real use, the README's scheduled-task example is the most concrete: create recurring work either by conversation or through the scheduled task UI, for instance a weekday 9 AM AI news digest. Skills are configured in SKILLs/skills.config.json, and the README counts 28 built-in skills covering web search, Word, spreadsheets, PowerPoint, PDF processing, Remotion video generation, browser automation, image and video generation, stock research, content writing, email, weather and skill creation. Build and test commands are listed separately: npm run build for the production renderer bundle, npm run compile:electron for the main and preload TypeScript build, npm test for the Vitest entry used by CI, and npm run lint across src, which the README warns may surface existing legacy debt.

Where LobsterAI gets in your way

The source-build story is the weakest documented area. The first run depends on a sibling ../openclaw checkout that the README never tells you how to obtain, and package.json pins openclaw to version v2026.8.1 from https://github.com/openclaw/openclaw.git. If that checkout is missing or the wrong revision, the documented first-run command has nothing to build against. Node is pinned to a single major line, >=24.15.0 <25, which will conflict with whatever else is on a developer machine.

Platform coverage is the second limit. The badges and installer links cover macOS and Windows only. There is no documented Linux desktop path and no headless mode, so a CI runner or a server cannot host this as described.

Approval friction is a design choice worth naming. The README states that LobsterAI asks for approval before sensitive actions such as file operations, terminal commands or network access. That is the right default for an agent holding your real desktop, but it makes long unattended runs harder than a sandboxed agent that never prompts. Finally, the README does not document rollback, downgrade or migration between releases, so an upgrade that changes the OpenClaw pin is not covered by any stated recovery procedure.

LobsterAI versus OpenClaw alone, and versus a sandboxed coding agent

The most useful comparison is against OpenClaw by itself, because LobsterAI is not a fork of the runtime but a client of it. OpenClaw provides agent execution and the gateway; LobsterAI adds the desktop shell: Electron lifecycle, SQLite session storage, approvals, artifact previews, multi-agent configuration, Expert Kits, MCP server management, scheduled tasks and IM bindings. Choosing between them is really choosing whether you want to assemble those pieces yourself. If you only need agent execution against a repository, the runtime alone is the smaller dependency; if you want a window that renders generated HTML, SVG, video, Mermaid diagrams and documents, and a phone channel to trigger work, that is the layer LobsterAI contributes.

The second comparison is against sandboxed coding agents that run in a container. Those trade access for safety: the agent cannot touch your real filesystem or your logged-in browser session, and in exchange it cannot build the local inventory system or check your ads dashboard that the README's prompts describe. LobsterAI takes the opposite position, running tools on the real desktop with an approval gate as the control. Neither is strictly better; they answer different questions about trust.

Licence, release cadence and what an upgrade actually costs

The project is MIT licensed, and package.json marks the package private, which is consistent with a desktop application rather than a published library. MIT is permissive, so redistribution and modification are allowed under its terms; this is a factual note about the licence identifier, not legal advice, and packaging concerns such as code signing are separate. The .env.example file shows what a maintainer needs to produce signed builds: APPLE_ID, APPLE_APP_SPECIFIC_PASSWORD and APPLE_TEAM_ID for macOS notarization, and YD_SIGN_SERVICE_URL, YD_SIGN_APP_KEY, YD_SIGN_APP_SECRET and YD_SIGN_USERNAME for Windows signing. The file states plainly that without the Windows values, builds are produced unsigned and are dev-only. Anyone forking to ship their own installer inherits that signing work.

Upgrade cost is driven by the OpenClaw pin rather than by LobsterAI's own version number. package.json pins openclaw to v2026.8.1 and lists provider plugins at matching versions, including @openclaw/qwen-provider, @openclaw/deepseek-provider, @openclaw/moonshot-provider, @openclaw/qianfan-provider, @openclaw/stepfun-provider, @openclaw/zai-provider, @openclaw/xiaomi-provider and @openclaw/volcengine-provider, all at 2026.8.1. Connector plugins move on their own schedule: @dingtalk-real-ai/dingtalk-connector at 0.8.26, @larksuite/openclaw-lark at 2026.7.16, @tencent-connect/openclaw-qqbot at 2.0.1. The last push to the repository was on 2026-09-09, and the most recent releases are 2026.9.4, 2026.9.3 and 2026.8.28. That is a frequent release rhythm, which cuts both ways: fixes arrive quickly, and so does the need to re-verify that a new LobsterAI build still matches the runtime and plugin versions it expects.

Multi-agent roles and IM channels: what the bindings actually change

Two features deserve separate scrutiny because they alter how the tool is used rather than how it is installed. Multi-agent workflows let you create Agents with their own identity, model choice, skills, working directory, enabled state and IM bindings, keeping a Main Agent for general work and specialized Agents for repeatable roles. Expert Kits add a second selection axis: scenario-oriented kits that package capability selections and references, chosen independently from direct skills, so a workflow can combine curated kits with individual tools. The practical effect is that a scheduled digest agent and a document-screening agent can hold different models and different working directories without sharing state.

IM remote control is the other differentiator, and it is broader than a single chat app. The README lists WeChat, WeCom, DingTalk, Feishu/Lark, QQ, Telegram, Discord, NetEase IM, NetEase Bee, POPO and email as reachable channels, and notes that multi-instance platforms can bind different accounts or channels to different Agents. That is a real operational difference from agents that only accept input from a local terminal. It also multiplies the surface you have to reason about: each connected channel is a way to trigger commands on your desktop, which is exactly why the approval gate exists.

Editorial conclusion

Adopt LobsterAI if you want an agent that touches your real desktop files, terminal and browser, and you are willing to run the OpenClaw runtime beside it. Skip it if you need a headless Linux deployment, a stable plugin API, or a source build on a Node version outside >=24.15.0 <25. Before committing, confirm you can build the pinned OpenClaw runtime from the sibling ../openclaw checkout, because npm run electron:dev:openclaw is the only documented first-run path.

Frequently asked questions

Official sources

1. License: MIT
2. netease-youdao/LobsterAI on GitHub
3. Project website
4. README
5. Releases