Lime - Full-Stack Desktop AI Agent
An open-source full-stack desktop AI agent that can actually finish the work, spanning coding, files, terminals, tools, research, content, and multimodal workflows.
- Source repo
- limecloud/lime
- Stars
- ★ 1.5k
- Last updated
- today
- Primary language
- TypeScript
- FA score
- 44/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer needs to fix a bug in a repository; Lime reads relevant files, traces the call path, implements changes, and runs focused tests.
- Not a fit if
- Teams needing a Linux desktop build; Linux builds are paused
- Teams needing closed-source commercial integration free of GPLv3 terms
- Source review
- 44/100 · Major gaps
What does this agent do, and when should you use it?
Lime is an open-source full-stack desktop AI agent for users and teams worldwide, bringing the agent loop, filesystem, terminal processes, code changes, tool calls, MCP, Skills, multimodal input and output, model routing, and multi-agent collaboration into one traceable task chain. It belongs to the same category of hands-on agents as Claude Code, WorkBuddy, and Codex, while emphasizing a desktop GUI, visual workspace, configurable providers, and mixed engineering, research, and content workflows. Built on Electron and a Rust App Server, Lime supports macOS and Windows and is licensed under GPLv3. It allows users to configure their own providers and models, extend the agent via MCP and Skills, and execute and review tasks in a unified workspace.
Lime performs the following operations: it understands goals, repositories, files, history, and constraints before proposing an executable plan; reads and writes files, searches, applies patches, runs terminal commands, tests, and calls tools within granted permissions; projects work as Thread, Turn, Item, and reusable artifacts, enabling pause, review, restore, and continuation; handles multimodal understanding (text, code, images, screenshots, audio, video, PDFs, tables) and generation (images, audio, video, documents, charts); discovers and calls external tools or local Skills via MCP; supports multi-agent collaboration, preserving shared context and permissions; allows per-task model selection and manages capability catalogs, credentials, routing, retries, and failure boundaries.
- A developer needs to fix a bug in a repository; Lime reads relevant files, traces the call path, implements changes, and runs focused tests.
- A full-stack team needs to ship a feature spanning frontend, App Server, Rust runtime, protocol, and tests; Lime splits and executes in dependency order.
- A researcher needs to turn web pages, notes, screenshots, and meeting records into a report or launch draft.
- A content creator needs to generate publish-ready output from multimodal material like audio, video, and PDFs.
- A team wants to encode a recurring check or release step as a reusable Skill for the agent to run via MCP.
- A coordinator needs to delegate research, implementation, testing, and documentation to different agents, collecting results in a main thread.
How do you install or deploy this agent?
Download the installer from Releases: macOS users can get the .dmg package or install via Homebrew (brew tap aiclientproxy/tap && brew install --cask lime); Windows users can download Lime_*_x64-setup.exe. Currently only macOS and Windows builds are published; Linux desktop builds are paused.
How do you use this agent?
- Open Lime, configure a Provider, and test a model connection. 2. Choose a workspace or project directory and confirm file and terminal permissions. 3. Create an Agent Thread with a goal, constraints, and acceptance criteria. 4. Ask for a plan first, then approve file changes, commands, or external tool calls as needed. 5. Inspect the diff, test results, and artifacts; continue the task or close it.
What are this agent's strengths and limitations?
- Desktop GUI and visual workspace differentiate it from CLI tools like Claude Code
- Provider-agnostic design avoids lock-in to a single model service
- Integrates MCP and Skills, allowing reusable procedures as execution units
- Native multimodal support for diverse input and output formats in one task
- Multi-agent collaboration with parallel subtasks and state preservation
- Limited platform support: only macOS and Windows; Linux builds paused
- Installers may be unsigned, triggering Windows SmartScreen warnings
- GPLv3 license may constrain commercial integration
- No bundled AI models, relying on user-configured third-party providers, with potential data upload
- Project maturity uncertain, with limited documentation and community support
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Lime - Full-Stack Desktop AI Agent This agent | 44 · Major gaps | Desktop appFree + model costs | ★ 1.5k | today | TypeScript | — |
| CodePilot — Multi-Model AI Agent Desktop | 47 · Major gaps | Desktop appFree + model costs | ★ 6.5k | 27d ago | TypeScript | OpenAI API · Claude API |
| iPolloWork | 44 · Major gaps | Desktop appFree + model costs | ★ 7.1k | 3d ago | TypeScript | — |
| Understanding AI Agents: Design Principles and Engineering Practice | 41 · Major gaps | CLIFree + model costs | ★ 55k | 19d ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows permission confirmation mechanism (README mentions approving high-risk actions), but no implementation details, so score 1. Data flow transparency: README states local storage by default, but no detailed data flow, score 1. Sensitive data handling: no explicit encryption or anonymization, score 1. Dependency security: lock file and frozen-lockfile in CI, but no vulnerability scanning mentioned, score 1. External effects: permission control mentioned, but no specific limits, score 1. Rollback: mentions resumable work, but no specific mechanism, score 1. Source attribution: no mention of source tracking, score 1.
Self-consistency: README and package.json descriptions align, score 2. Dependency availability: lock file present, but no availability guarantee, score 1. Failure messages: no error handling documentation, score 1.
Audience and scenarios: README clearly lists target users and scenarios, score 2. Capability boundaries: no explicit permission boundaries, score 1. Trigger precision: no trigger mechanism details, score 1. Environment fit: supports macOS and Windows, score 2.
Information architecture: README and docs directory present, score 2. Install notes: download and Homebrew instructions provided, score 2. Naming stability: no API stability statement, score 1. Examples and FAQ: FAQ and examples present, score 2. Known limitations: not explicitly listed, score 1. License: GPLv3 clearly stated, score 2. Versioning and changelog: RELEASE_NOTES file exists, but content not provided, score 1. Maintenance responsibility: no explicit maintainer, score 1.
Output usability: describes generating reports, scripts, etc., score 2. Marginal value: provides multimodal and multi-agent features, score 2. Cost benefit: no performance or cost data, score 1.
Claim traceability: README claims features, but no evidence, score 1. Cross-source corroboration: only README and package.json, score 1. Fact-inference separation: not explicitly distinguished, score 1.
- No specific permission implementation details, need further code review.
- No detailed data flow and sensitive data handling documentation.
- No dependency vulnerability scanning information.
- No known limitations and rollback mechanism specifics.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.162.0 Lime v1.162.0
Release notes excerpt
Lime v1.162.0 新功能 - 统一窗口页头与工作区标签，整合会话动作和右侧面板入口，支持分隔条拖动与键盘调整。 - 新增基于会话事实的摘要面板与固定偏好，让会话概览和工作区内容共用当前状态。 - 支持远端插件及逐技能启停，列表和预览共享账号偏好，保存后重新读取安装事实。 修复 - 修正 Responses 流式工具调用的身份关联与终态去重，避免将参数事件识别为额外调用。 - 修正工具历史截断、预算和读取结果投影，保留工具调用与输出的关联及原始记录。 - 上下文压缩改为由当前模型生成摘要，补齐取消、持久化及分叉历史的相关处理。 - 修正会话分叉导航的草稿作用域提交，避免父子会话草稿相互影响。 - 加强插件偏好操作的账号隔离、失败保留和刷新处理；需要 Apps 授权的技能保持受控禁用。 优化与重构 - 收敛远端插件目录、详情与安装事实的来源，保持本地运行时配置和远端账号偏好的领域边界。 - 以共享页头、标签和摘要实现替换旧任务中心标签与 Activity 面板，删除首页无消费的回调和派生状态。 - 将上下文压缩、历史归一化、Hook 生命周期等逻辑拆入现有领域模块。 测试与质量 - 补充工作区布局、摘要固定、插件偏好、工具调用身份与历史预算回归，以及相关 Electron 和公共 JSON-RPC 场景。 …View on GitHub ↗ -
v1.161.0 Lime v1.161.0
Release notes excerpt
Lime v1.161.0 新功能 - 工作台新增 PDF、DOCX、工作簿和演示文稿的原文件预览、位置批注与会话反馈；工作簿和演示文稿支持已实现范围内的编辑与保存。 - 扩展插件市场来源管理、远程插件获取、插件共享与访问范围配置，统一桌面、CLI 与 App Server 的插件合同。 - 新增 App Server 账号与云端请求接口，统一桌面登录、账号状态同步及退出流程。 修复 - 修复矮窗口中首页输入区和底部控件被裁切的问题。 - 退出应用时等待 App Server 异步停止，完善桌面进程的关闭流程。 - 修复专业文件字节摘要计算可能导致预览或批注初始化失败的问题。 - 修复工作区切换、文件预览与批注提交中的状态归属、迟到响应和文件版本冲突处理。 - 完善 MCP、实验功能与媒体服务设置的未保存更改保护，以及开发者和媒体设置的搜索定位。 - 调整 App Server 请求超时与诊断，完善插件刷新、安装身份及 Hooks 生命周期处理。 优化与重构 - 主导航采用折叠后仍保留的图标栏、上下文侧栏和共享标题栏；简化首页，移除旧推荐入口，保留 Banner 与独立输入区。 - 完善输入区工作树提交、执行范围交接和后台会话恢复状态同步。 …View on GitHub ↗ -
v1.160.0 Lime v1.160.0
Release notes excerpt
Lime v1.160.0 新功能 - 设置新增内容搜索，可定位网页搜索、记忆和外观等页面的具体字段与内部页签，并保留未保存更改保护。 - 插件目录、详情、安装和 CLI 入口统一使用市场与插件身份，Skills、MCP 和 Apps 复用共享配置及安装状态。 - 完善会话工作目录、服务档位与审批审核者在创建、恢复和分叉过程中的传递与持久化。 修复 - 修复动态工具宿主在连接断开与会话恢复后的状态清理和重新接管。 - 加强 Skills 根目录监听、插件配置保存与安装状态一致性，避免多入口产生不同事实源。 - 修复 macOS 公证已接受后遇到临时票据装订网络故障的发布恢复，并保留签名与公证失败诊断。 优化与重构 - CLI、TUI 与桌面继续收敛到同一 App Server 协议、运行时和会话投影，拆分连接、历史、恢复及工作目录职责。 - 直接替换旧插件协议和目录直装入口，移除旧 mock 与过期文档，不保留双轨实现。 - 拆分记忆与外观设置的展示组件，复用搜索、导航、草稿和配置 owner；同步五语言界面资源。 测试与质量 - 扩充插件市场、审批、服务档位、远程工作目录、动态工具和 Skills 的协议及运行时回归。 - 增加设置搜索、内部页签定位、未保存导航与五语言测试，并更新真实 Electron 场景。 …View on GitHub ↗