M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
L

LobsterAI

Community
Not a standard server
NetEase Youdao's open-source desktop AI agent for office work — an MCP client/host, not an MCP server itself

This is not a typical connect-and-use MCP server; read the source repository before using it.

How to install ↓ Compare GitHub source repository ↗
Category
Other
Stars
★ 6.2k Very popular
Runtime
Node.js >=24.15.0 <25 · Prebuilt binary
Credentials
No credential needed
License
MIT
Last commit
50FMRS · D

The source material describes LobsterAI as an open-source desktop AI agent application (an MCP client/host), not an MCP server that other clients connect to; no evidence of an MCP server manifest, tool list, transport, or install configuration was found, so this profile focuses on its actual functionality, permissions, and risks as a desktop automation tool.

Strongest · Documentation 14/20 Weakest · Reliability 5/20

Reliability
5/20
Security and permissions
10/20
Maintenance
13/20
Documentation
14/20
Setup experience
8/20
Why each score
Reliability 5/20
The core issue: this repository is fundamentally an Electron desktop application (built on the OpenClaw runtime). The README explicitly frames it as an MCP *client* — users configure external MCP servers inside it and LobsterAI syncs them into OpenClaw — not as a repository that itself exposes a set of MCP tools. No server manifest was provided (marked not cached), so there is no evidence of any tool declarations, an MCP init handshake, or a tool-list response to verify that declared tools match real behavior. The README references 'npm test' as 'the official Vitest entry used by CI' and lint scripts, but no actual CI workflow file or test coverage over key paths is shown, so it does not meet the static-calibration bar requiring real CI workflows plus committed tests. Per the rule, reliability is capped at 12 when this evidence is absent; combined with the object-type mismatch and total absence of MCP-specific evidence, a low score is warranted.
Security and permissions 10/20
The README (untrusted evidence, considered only as a claim) states renderer windows use context isolation, disabled Node integration, and sandboxing, that sensitive tool actions (file ops, terminal commands, network access) are permission-gated and logged, and that app data is stored locally in SQLite. None of these claims can be independently verified without source or manifest access. The product's attack surface is notably broad: it allows remote control of a real desktop (terminal commands, file operations) via numerous IM channels (WeChat, Feishu, DingTalk, Telegram, Discord, QQ, WeCom, etc.), and inadequate confirmation on such remote-trigger paths would be high-impact. No explicit red-line evidence was found (no malware, credential theft, fabricated identity/test results), but because the permission model, confirmation flow, and data boundaries cannot be independently verified from static material alone, a high score is not justified.
Maintenance 13/20
The repository has 5775 stars, 757 open issues, a clear MIT license, and is not archived — signs of real activity and an active user base; the README's contribution guidance (issue/PR templates) also indicates some baseline governance. However, there is no direct evidence of recent commit cadence, release/versioning history, or a security-response channel (e.g., SECURITY.md), and the relatively high open-issue count may suggest triage strain. Overall this sits in the upper-middle range: active, but with governance/versioning evidence gaps.
Documentation 14/20
The README is well-structured, covering installation (source build and downloadable installers), development commands, a project map, packaging instructions, and a dedicated 'Security And Data' section, plus real-world usage prompts — solid layered documentation. However, since this object is not exposed as an MCP server with tools, there is no documentation at all of tool parameters, invocation examples, error handling, rate limits, or cost, and no troubleshooting content from that angle — a documentation gap directly relevant to FMRS scoring criteria.
Setup experience 8/20
This is not a server that can be connected to a mainstream MCP client in a few steps via npx/uvx; it is a full Electron desktop application requiring a pinned Node.js version (>=24.15.0 <25), npm install, building the OpenClaw runtime (with multiple environment-variable branches for force-rebuild, custom source paths, etc.), and separate packaging pipelines per platform (macOS/Windows/Linux). The process involves many steps and a substantial desktop build toolchain, and no mcpServers configuration JSON example is given showing how to register it as an MCP endpoint in a client like Claude Desktop. Per static calibration, setup cannot exceed 15 without verifiable execution evidence, and the process complexity further lowers the score.

Static review · not runListed 2026-08-06

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or codeUses the networkControls a browser

Best for

  • Users who want a desktop-grade AI agent that can directly operate local files, terminal, and browser workflows
  • Users who want to remotely trigger desktop Agent tasks via WeChat, Feishu, DingTalk, Telegram, or similar IM channels
  • Users needing office automation for documents, spreadsheets, slides, video, and web research
  • Developers comfortable with Electron/React/TypeScript who want to build from source or use official installers

Not for

  • Users who need an MCP server that a standard MCP client (e.g. Claude Desktop) can connect to directly — no such server manifest or tool definitions are provided in this repository
  • Users needing a headless or server-side deployment — this project only ships as a macOS/Windows desktop application
  • Users who want a narrow, single-purpose tool integration without granting broad file, terminal, and network permissions

Required permissions

  • Local file system read/write
  • Terminal command execution
  • Browser automation
  • Network access
  • IM channel binding and messaging (WeChat, WeCom, DingTalk, Feishu/Lark, QQ, Telegram, Discord, NetEase IM, NetEase Bee, POPO, email)
  • Local SQLite storage (lobsterai.sqlite under Electron userData)

Risks and side effects

  • The app requires broad, sensitive permissions (file operations, terminal, network); the README notes sensitive actions are permission-gated and logged, but users must still review approval prompts carefully to avoid data loss or exposure
  • Remote control via IM channels means commands can trigger desktop actions from linked messaging accounts — compromise of a bound account/channel could lead to unauthorized local execution
  • Packaging bundles the OpenClaw runtime and, on Windows, a portable Python runtime, expanding the local attack surface
  • This project is not maintained by the MCP protocol owner and is not a conventional MCP server; anyone wanting standard MCP-protocol integration should verify exposed behavior directly, as the source material gives no such details

Setup

Before you start

Runtime：Node.js >=24.15.0 <25 · Prebuilt binary

Other optional settings (8)
OPENCLAW_SRC optional Dev-only path to a local OpenClaw source checkout, used with npm run electron:dev:openclaw.
OPENCLAW_FORCE_BUILD optional Set to 1 to force a rebuild of the OpenClaw runtime.
OPENCLAW_SKIP_ENSURE optional Set to 1 to keep a local OpenClaw checkout on its current branch/tag.
LOBSTERAI_PORTABLE_PYTHON_ARCHIVE optional Path to a portable Python archive for offline/private packaging.
LOBSTERAI_PORTABLE_PYTHON_URL optional URL to a portable Python download for offline/private packaging.
LOBSTERAI_WINDOWS_EMBED_PYTHON_VERSION optional Pins the embedded Python version for Windows packaging.
LOBSTERAI_WINDOWS_EMBED_PYTHON_URL optional Custom download URL for the embedded Python runtime in Windows packaging.
LOBSTERAI_WINDOWS_GET_PIP_URL optional Custom download URL for the get-pip script in Windows packaging.

The source material does not describe an install configuration for connecting to this project as an MCP server from another client. The repository itself is installed as a desktop application: download the macOS/Windows installer from the official site or GitHub Releases, or clone the repo and run npm install with Node.js >=24.15.0 <25; first development run uses npm run electron:dev:openclaw, subsequent runs use npm run electron:dev (renderer dev server at http://localhost:5175).

Check that it works

Install LobsterAI from the official website or GitHub Releases, launch the desktop app, start a Cowork session, and send a simple request (e.g., asking the agent to list local files) to confirm it streams progress and executes tools.

Troubleshooting

  1. Verify Node.js version meets >=24.15.0 <25
  2. First run requires `npm run electron:dev:openclaw` to build/pin the OpenClaw runtime; subsequent runs can use `npm run electron:dev`
  3. Force a runtime rebuild with the `OPENCLAW_FORCE_BUILD=1` environment variable if needed
  4. Point to a custom OpenClaw checkout via the `OPENCLAW_SRC` environment variable
  5. Check that permission prompts for sensitive actions (file, terminal, network) are being approved correctly
  6. Inspect gateway logs under `userData/openclaw` for runtime issues
  7. In source-build mode, confirm the renderer dev server is running at http://localhost:5175

Things to try

Once connected, you can ask your AI assistant things like:

  • Use product-growth.xlsx to build a visual dashboard and summarize the main growth drivers.
  • Research the AI Agent market and turn the findings into a presentation.
  • Open the ads dashboard every morning, check spend and conversion anomalies, and summarize likely causes.
  • Turn the resumes in this folder into a screening sheet and shortlist the strongest candidates against the JD.

Use cases

Turn an Excel-based inventory workflow into a local inventory/sales system
Build a visual dashboard from a local data file and summarize growth drivers
Research a market and generate a presentation deck
Open an ads dashboard daily, check spend/conversion anomalies, and summarize causes
Turn a folder of resumes into a screening sheet against a job description
Set up scheduled tasks like a daily news digest at a fixed time

Overview

LobsterAI is an open-source desktop Agent application from NetEase Youdao (Electron-based, macOS and Windows) that operates in a real desktop environment — local files, terminal commands, browser workflows, documents, spreadsheets, slides — and can be remotely controlled via IM channels such as WeChat, WeCom, Feishu/Lark, DingTalk, and Telegram. Its product/session layer is called Cowork, running on the OpenClaw execution runtime underneath. The README states LobsterAI can connect to external Model Context Protocol (MCP) servers as tools/data sources, storing user-configured servers locally and syncing enabled ones into OpenClaw — meaning LobsterAI itself acts as an MCP client/host, not as a server that exposes MCP tools. No MCP server manifest, tool list, transport, or install configuration for consuming LobsterAI as an MCP server is present in the source material, so those fields are left empty rather than guessed.

Similar servers

Telegram MCP Server 70 · B

Telegram MCP server powered by Telethon to let MCP clients read chats, manage groups, and send/modify messages, media, contacts, and settings.

★ 2.0k · Tools 76 Compare with this →
Compare with other MCP servers

Source revision da88493b5a8d Data synced 2026-11-04 Read the FMRS scoring method