CodePilot — Multi-Model AI Agent Desktop
Connect any AI provider, extend with MCP & skills, control from your phone, and let your assistant learn your workflow.
- Source repo
- op7418/CodePilot
- Stars
- ★ 6.5k
- Last updated
- 27d ago
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 47/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers who juggle multiple AI providers and want a single desktop client with consistent UI.
- Not a fit if
- Companies unwilling to buy a separate commercial license
- Windows-only organizations that cannot accept unsigned installers
- Source review
- 47/100 · Major gaps
What does this agent do, and when should you use it?
CodePilot is a multi-model AI agent desktop client built with Electron + Next.js. It connects to 17+ AI providers out of the box, including Anthropic, OpenRouter, AWS Bedrock, Google Vertex AI, Zhipu GLM, Kimi, Moonshot, MiniMax, Volcengine Ark, Xiaomi MiMo, Aliyun Bailian, Ollama, LiteLLM, and any Anthropic- or OpenAI-compatible endpoint. The app offers three interaction modes (Code, Plan, Ask), session control (pause, resume, rewind to checkpoint), split-screen dual sessions, and permission control. Beyond coding, it includes an Assistant Workspace for persona files, persistent memory, onboardings, and daily check-ins; generative UI for AI-created dashboards; remote bridge to Telegram, Feishu, Discord, QQ, and WeChat; MCP server support (stdio/sse/http) with runtime monitoring; skills from the skills.sh marketplace; image generation; and a task scheduler. All data is stored locally in SQLite (WAL mode), ensuring privacy and ownership.
CodePilot provides a desktop interface to initiate multi-turn conversations with any configured provider. Users can select a working directory, mode, and model; attach files and images; switch models mid-conversation; and control permission levels (default/full access or per-action approval). It can pause, resume, and rewind sessions to any checkpoint, archive sessions, and run two sessions side by side. MCP servers (stdio/sse/http) can be added with runtime monitoring, and custom skills can be defined or installed from the skills.sh marketplace. The remote bridge allows sending messages from Telegram, Feishu, Discord, QQ, or WeChat to trigger actions on the desktop. Generative UI enables AI-created interactive dashboards and widgets rendered live. Image generation supports batch tasks and a gallery. The task scheduler runs recurring tasks using cron expressions or intervals. CodePilot also imports Claude Code CLI .jsonl session history, tracks token usage and costs with daily charts, and provides a file browser and Git panel.
- Developers who juggle multiple AI providers and want a single desktop client with consistent UI.
- Users who want to control their AI assistant from their phone via Telegram, WeChat, or other bridge channels.
- Power users who need to extend the agent with custom MCP servers and skills for specific workflows.
- Data analysts or managers who want AI-generated dashboards and visual widgets inside the app.
- Teams that need to monitor token usage and costs with built-in analytics and daily charts.
- Privacy-conscious users who prefer local storage (SQLite) and self-hosted models like Ollama.
How do you install or deploy this agent?
Option A: Download a release\nVisit the Releases page and download the installer for your platform (macOS .dmg, Windows .exe, Linux AppImage/deb/rpm), then launch the app.\n\nOption B: Build from source\n``bash\ngit clone https://github.com/op7418/CodePilot.git\ncd CodePilot\nnpm install\nnpm run electron:dev # full desktop app\n``\nRequires Node.js 18+ and npm 9+.
How do you use this agent?
- After launching, go to Settings > Providers and add credentials for any supported provider (e.g., Anthropic, OpenRouter, Ollama).\n2. Start a conversation: choose a working directory, mode (Code/Plan/Ask), and model.\n3. (Optional) In Settings > Assistant, set up a workspace and enable onboarding; CodePilot creates
soul.md,user.md,claude.md, andmemory.mdat the workspace root.\n4. (Optional) Add MCP servers via the MCP page in the sidebar; manage custom skills on the Skills page.\n5. (Optional) For advanced features like file editing and terminal commands, install the Claude Code CLI:npm install -g @anthropic-ai/claude-code.\n6. (Optional) Configure remote control via the Bridge page for Telegram, Feishu, Discord, QQ, or WeChat.
What are this agent's strengths and limitations?
- Supports 17+ providers, including local models (Ollama) and Chinese AI services, offering unmatched flexibility.
- Built-in MCP support (stdio/sse/http) with runtime monitoring, enabling easy tool integration.
- Remote bridge lets you control your desktop assistant from your phone via popular messaging apps.
- Includes advanced features like generative UI, image generation, and task scheduling, going beyond simple chat.
- Session management with pause/resume/rewind and split-screen dual sessions enhances productivity.
- License is BSL-1.1 with commercial use requiring a separate license; code won't become Apache 2.0 until 2029.
- Release builds are unsigned/unnotarized, causing Gatekeeper or SmartScreen warnings.
- Advanced features (file editing, terminal commands) depend on the Claude Code CLI installation.
- Each bridge channel requires its own bot token and configuration.
- Some cloud providers (Bedrock, Vertex) need additional IAM setup beyond API keys.
How does this agent compare with similar options?
Compared to Claude Code CLI, CodePilot offers a graphical interface, multi-provider support, and built-in extensions like MCP and remote bridge, but advanced coding features still require the CLI. It positions itself as a general-purpose AI desktop agent rather than a coding-only tool, directly competing with vendor-specific desktop clients but remaining provider-agnostic.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| CodePilot — Multi-Model AI Agent Desktop This agent | 47 · Major gaps | Desktop appFree + model costs | ★ 6.5k | 27d ago | TypeScript | OpenAI API · Claude API |
| Kun Workspace | 49 · Major gaps | Desktop appFree + model costs | ★ 6.3k | 1d ago | TypeScript | ChatGPT · Codex · OpenAI API · Claude API |
| Lime - Full-Stack Desktop AI Agent | 44 · Major gaps | Desktop appFree + model costs | ★ 1.5k | today | TypeScript | — |
| PI-Desktop | 68 · Some gaps | Desktop appFree + model costs | ★ 7k | today | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows permission controls (Default/Full Access, per-action approval) and user confirmation mechanisms (e.g., CLI installation requires user action), but data flow and sensitive data handling are not detailed. Dependency security has overrides but no audit. External effects (e.g., remote bridge) are mentioned but not deep. Rollback has session rewind but no system-level rollback. Source attribution has author info but unverified.
Self-consistency is good, README aligns with code structure, but dependency availability unverified (e.g., no check if dependencies are fetchable). Failure messages have structured errors (e.g., NO_CREDENTIALS), but not comprehensive.
Target audience clear (developers, daily users), scenarios diverse (coding, assistant, remote control). Capability boundaries described (e.g., CLI optional), but trigger precision (e.g., permission control details) insufficient. Environment fit has platform notes, but not all environments covered.
Information architecture clear (README, doc links), install notes detailed (download, build). Naming stability not explicit (version exists but no naming convention). Examples and FAQ provided, but known limitations not listed. License clear (BSL-1.1), versioning changelog not provided. Maintenance responsibility has author info but not explicit.
Output usability high (multi-platform, multi-format), marginal value high (multi-model, MCP, remote control), but cost-benefit not quantified (e.g., pricing, resource usage).
Claims traceable (README, code), but cross-source corroboration insufficient (single source). Fact-inference separation unclear (e.g., performance claims without benchmarks).
- Unverified publisher identity, proceed with caution.
- Dependency security lacks audit, recommend checking for vulnerabilities.
- Remote bridge feature may introduce security risks, pay attention to permission controls.
- License is BSL-1.1, commercial use requires separate license.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.67.17 CodePilot v0.67.17
Release notes excerpt
CodePilot v0.67.17 > 新增可查看来源、更正和忘记的助理记忆管理，减少辅助功能失败后的重复请求，并修复 Gemini 普通聊天报错。 新增功能 - 助理记忆管理 — 在设置的助理工作区中手动保存、查看来源、更正或忘记记忆；手动管理不调用模型。 - 自动提取状态 — 查看成功对话后的记忆提取进度、失败原因和可重试时间；同一批任务连续生成失败达到3次后停止自动重试。 - 快捷建议重试 — 动态建议不可用时显示原因，提供重试按钮和真实冷却截止时间。 修复问题 - 修复 Gemini 在启用视频工具后，连普通问候也可能因工具参数声明不兼容而报错的问题；视频时长仍严格校验为6秒或10秒。 - 修复只为 Claude Code 配置凭据时，后台记忆提取和快捷建议仍误尝试直接请求、重复产生错误的问题。 - 修复限流及普通请求错误可能永久阻断辅助功能的问题；凭据错误等待配置修复，其他错误冷却后可重试。 - 修复辅助配置文件不可用时可能打断聊天的问题；记忆搜索的可选智能排序失败时保留基础搜索结果并说明原因。 - 修复只在记忆正文出现的信息可能无法被搜索到，以及已忘记内容可能被旧索引再次带回的问题。 - 修复旧助理会话身份识别与界面不一致，以及已选择助理目录的远程会话无法参与自动提取的问题。 …View on GitHub ↗ -
v0.67.16 CodePilot v0.67.16
Release notes excerpt
CodePilot v0.67.16 > 新增 Gemini AI Studio 的 Native 接入，修复部分模型续聊时的路由不匹配，并改进消息保存失败提示。 新增功能 - Gemini AI Studio — 在官方 API 服务中添加 AI Studio API Key，即可在 Native（CodePilot）中选择 Gemini 3.8 Flash；本次不开放到 Claude Code 或 Codex。 - Gemini 思考深度 — 支持低、中、高三档思考深度，默认中档；支持工具调用历史的保存与同模型续聊。 修复问题 - 修复部分第三方模型首轮回复后，续聊出现模型路由不匹配的问题；覆盖 Windows 用户反馈的同类场景。 - 修复 Native 长对话压缩后重复带入旧消息，以及摘要可能产生连续用户消息的问题，保留原有图片附件。 - 回复达到输出长度上限时保留已生成正文并显示提示，避免误认为回答已经完整结束。 - 消息未确认保存时提供明确提示，保留当前正文，避免重试或后台刷新覆盖尚未保存的内容。 - 保存完成后的后台处理不再延迟前台回复结束；快捷建议失败后短暂退避，减少重复请求。 优化改进 - TokenDance 添加卡片移到授权登录区域，方便找到浏览器授权入口。 …View on GitHub ↗ -
v0.67.15 CodePilot v0.67.15
Release notes excerpt
CodePilot v0.67.15 > 加入 TokenDance：一次连接即可在受支持的 Runtime 中选择精选模型，并提供授权、模型发现和凭据恢复入口。 新增功能 - TokenDance 接入 — 在设置的服务商页面添加 TokenDance，可通过浏览器授权、一次性授权码或手动 API Key 连接。 - 六款默认精选模型 — 默认启用 GLM 5.3、Kimi K3、MiniMax M3、DeepSeek V4 Flash、DeepSeek V4 Pro 与 GLM 5.3 Flash；模型刷新保留用户手动隐藏、启用和编辑的选择。 - 同一连接跨 Runtime 使用 — CodePilot 与 Codex 使用聊天协议，Claude Code 使用模型明确支持的 Messages 协议，无需重复授权。Kimi K3 当前仅在 CodePilot 与 Codex 中显示。 - 授权与额度恢复提示 — 区分余额不足、Key 不可用与周期额度限制，提供充值或重新授权的明确指引。 优化改进 - 服务商、模型页与聊天选择器统一使用 TokenDance 官方 Logo，并显示多协议能力和对应 Runtime 的模型筛选结果。 - 授权支持取消与超时，重新授权沿用原连接；Key 仅在服务端加密保存，不返回页面。 …View on GitHub ↗