LobsterAI
An open-source desktop assistant for executing office work in your real local environment.
Per-dimension scores and reasoning
Evidence shows: README claims approval before sensitive actions, but no implementation details; permission gating and logging are claimed but not shown in code; many dependencies but no security audit or vulnerability scanning evidence; external effects (file ops, network) mentioned but not detailed; rollback not mentioned; source attribution has author info but publisher unverified. Deductions: lack of concrete implementation evidence, rollback missing.
Evidence shows: README and package.json are largely consistent in feature description, but some inconsistencies (e.g., README mentions 28 skills but not all listed); many dependencies but no availability guarantee; failure messages not mentioned. Deductions: dependency availability unverified, failure messages missing.
Evidence shows: README describes multiple scenarios (data analysis, slides, docs, etc.) with clear audience; capability boundaries mentioned (e.g., permission gating) but not detailed; trigger precision not mentioned; environment fit includes macOS and Windows but not Linux. Deductions: trigger precision missing, capability boundaries not detailed.
Evidence shows: README structure clear with project map; install notes detailed; naming stable (LobsterAI); examples and FAQ (community support) present; known limitations not mentioned; MIT license clear; version number exists but no changelog; maintenance responsibility has author info but unverified. Deductions: known limitations missing, changelog missing.
Evidence shows: output usability described (e.g., artifact preview) but not verified; marginal value described (multiple scenarios) but not quantified; cost-benefit not mentioned. Deductions: cost-benefit missing.
Evidence shows: some claims in README are supported by code (e.g., security measures), but some not verified; cross-source corroboration limited (only repo files); fact-inference separation unclear. Deductions: insufficient cross-source corroboration, unclear fact-inference separation.
- Publisher identity unverified, proceed with caution.
- Many dependencies, including Git dependencies and private registry, supply chain risk needs assessment.
- Rollback mechanism missing, may not recover after failed operations.
- Known limitations not disclosed, potential hidden risks.
What does this agent do, and when should you use it?
LobsterAI is an Electron desktop application maintained by NetEase Youdao for long-running work across local files, documents, spreadsheets, slides, browser workflows, and project workspaces. Its Cowork session layer owns local persistence, permissions, UI state, artifacts, agents, memory, and instant-messaging bindings, while OpenClaw provides the underlying execution runtime and gateway. The React renderer exposes sessions, agents, skills, MCP, scheduled tasks, and IM settings; the Electron main process handles IPC, SQLite, authentication, logging, and runtime management. The app can preview generated HTML, SVG, images, video, Mermaid diagrams, code, Markdown, text, and office-document artifacts. It fits users or teams that want an agent to operate on their actual desktop environment with approval for sensitive actions.
A user submits work in a Cowork session, where the application streams progress and tool output while retaining session history; it requests approval for sensitive file operations, terminal commands, and network access. openclawConfigSync renders locally configured providers, models, Agents, IM bindings, skills, MCP servers, and workspace instructions into OpenClaw configuration. openclawEngineManager manages the gateway process, runtime state, ports, logs, restart, and repair, while openclawRuntimeAdapter translates gateway events into Cowork stream events. The application stores sessions, messages, configuration, agents, and memory metadata in SQLite, and retains workspace memory in files including MEMORY.md, USER.md, SOUL.md, and daily notes. It also manages scheduled work and can connect the desktop agent to WeChat, WeCom, DingTalk, Feishu/Lark, QQ, Telegram, Discord, email, and other listed IM channels.
- A small business still tracking purchases and sales in Excel can ask it to build a local inventory system that calculates stock and profit and opens in a browser.
- An operations analyst working with a local
product-growth.xlsxfile can produce a visual dashboard and a summary of growth drivers. - A business user preparing an AI-agent market briefing can request research and a presentation from the findings.
- A marketer who checks an ads dashboard each morning can automate browser checks and receive a summary of spend or conversion anomalies.
- A recruiter screening resumes in a local folder can turn them into a screening sheet and shortlist candidates against a job description.
- A user away from their computer can send work to a specifically bound desktop agent through a supported IM channel.
What are this agent's strengths and limitations?
- Cowork and OpenClaw have a clear split: the desktop application retains local state, permissions, memory, and artifacts while the runtime handles agent execution and gateway behavior.
- It spans local files, terminal commands, browser workflows, office documents, spreadsheets, and slides, with in-app handling for varied generated artifacts.
- Sensitive actions are permission-gated and logged; renderer windows use context isolation, disabled Node integration, and sandboxing.
- Specialized Agents can have separate model choices, skills, working directories, and IM bindings, alongside scheduled tasks and MCP server management.
- Source development explicitly requires Node.js 24.15.x and npm, and the first run must build or ensure the pinned OpenClaw runtime.
- The supplied material does not specify supported model providers, authentication credentials, or pricing, so usable models and operating cost require separate validation.
- It can access local files, terminal commands, and network workflows; despite approval controls, adopters must evaluate permission scope and workspace data handling.
- Official end-user installation is documented for macOS and Windows; Linux has a packaging command but no end-user installation path in the supplied material.
How do you install or deploy this agent?
Desktop installers for macOS and Windows are available through the official site or GitHub Releases. Running from source requires Node.js >=24.15.0 <25 and npm:
git clone https://github.com/netease-youdao/LobsterAI.git
cd LobsterAI
npm installFor the first development run, use: npm run electron:dev:openclaw. Once the pinned OpenClaw runtime exists locally, use: npm run electron:dev. The supplied material does not document model-provider credential setup or the credentials required for a first working conversation.
How do you use this agent?
After launch, enter a task concerning local projects, files, or browser work in a Cowork session, and approve sensitive file, terminal, or network actions when prompted. Create specialized Agents in the agent settings with their own identity, model choice, skills, working directory, enabled state, and IM binding; use MCP settings to add external MCP servers, and create recurring work through conversation or the scheduled-task UI. In development mode, the renderer server runs at http://localhost:5175.
FAQ
Which actions require approval?
Where does it store data?
lobsterai.sqlite under Electron userData. OpenClaw state, workspace memory, generated configuration, and gateway logs live under userData/openclaw.