MyAgents
An open-source, local-first desktop Agent workbench that unifies conversations, workspaces, files, tools, models, and task scheduling into one desktop system for real workflows.
- Source repo
- hAcKlyc/MyAgents
- Stars
- ★ 579
- Last updated
- 1d ago
- License
- AGPL-3.0
- Primary language
- TypeScript
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer has the Agent work inside a real project folder — reading code, running terminal commands, and verifying results in a split panel instead of a detached chat box
- Not a fit if
- Users who want zero-config use without supplying API keys
- Businesses needing closed-source modification or OEM distribution without AGPL obligations
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
MyAgents (GitHub: hAcKlyc/MyAgents) is an open-source desktop workspace for personal AI Agents, built with Tauri v2 + Rust + React 19 and licensed under AGPL-3.0. Rather than being a chat window, it puts multi-tab sessions, workspace file trees, an embedded terminal, an embedded browser, a task center, and local full-text search into a single application. Execution runs through a Node.js v24 Sidecar with the built-in Claude Agent SDK, and a Lab mode lets sessions be driven by Claude Code CLI, OpenAI Codex CLI, or Google Gemini CLI instead. Data is local-first by default: conversations, tasks, config, and generated artifacts stay on your machine. It suits advanced users who want to configure their own model providers, skills, and automations, and want AI deeply integrated with local files and the terminal.
MyAgents organizes Agents around workspaces: each Tab is bound to a real folder, sessions reference files with @, invoke Skills with /, and call MCP tools (STDIO/HTTP/SSE), while a split-panel embedded terminal (xterm.js + portable-pty) and embedded browser (Tauri multi-Webview) let you execute and verify in place. The Node.js Sidecar embeds the Claude Agent SDK by default, with Claude Code CLI / OpenAI Codex CLI / Google Gemini CLI available as external runtimes in the Lab. Model providers are configurable in-app: Anthropic subscription or API, plus DeepSeek, Moonshot, Zhipu, MiniMax, Google Gemini, Volcengine, SiliconFlow, ZenMux, OpenRouter, Xiaomi MiMo, Alibaba Cloud Bailian, and more. The Ideas + Task Center turns rough thoughts into executable tasks, supporting one-shot and recurring tasks via Cron expressions, optional local-command sensors that wake the AI only on condition match, execution logs, and failure recovery. IM integration ships with Telegram and DingTalk bots, with other platforms available through OpenClaw plugins. Local full-text search (Tantivy + jieba) indexes session history and workspace files. Architecturally, Session:Sidecar is 1:1, control-plane HTTP/SSE goes through a Rust proxy, and workspace file IO goes through Tauri/Rust.
- A developer has the Agent work inside a real project folder — reading code, running terminal commands, and verifying results in a split panel instead of a detached chat box
- An individual captures scattered thoughts in the Ideas center, aligns on goals with the AI, and converts them into scheduled, state-tracked tasks
- A heavy Agent user runs several independent sessions in parallel via Chrome-style tabs, each with its own Sidecar
- A user who wants a custom model mix (e.g., Anthropic subscription + DeepSeek + OpenRouter) without being locked to one vendor
- A user who wants the Agent to handle messages via a Telegram or DingTalk bot, or execute tasks automatically on a Cron schedule
- A developer familiar with Claude Code CLI / Codex CLI / Gemini CLI who wants one desktop interface to switch between Agent runtimes
How do you install or deploy this agent?
End users can download the app from https://myagents.io; it supports macOS 13.0+ (Apple Silicon and Intel) and Windows 10+. For a developer build you need Node.js >=22.0.0 (Node 24 recommended; the repo declares [email protected]), Rust installed via rustup (toolchain pinned by rust-toolchain.toml), and for cold native-inference builds CMake 3.28+ plus a platform C/C++ toolchain (macOS additionally needs Git, Python 3.10+, and Apple Clang).
macOS / Linux:
git clone https://github.com/hAcKlyc/MyAgents.git
cd MyAgents
./setup.sh
./start_dev.shWindows:
powershell
git clone https://github.com/hAcKlyc/MyAgents.git
cd MyAgents
.\setup_windows.ps1
.\build_windows.ps1On first run you must configure at least one model provider (e.g., an Anthropic API key or subscription) in the in-app Model Providers page, otherwise the Agent cannot execute anything.
How do you use this agent?
- Launch the app and open or create a workspace (each Agent binds to a real folder). 2. Configure models in the Model Providers page (Anthropic subscription/API, or DeepSeek, Moonshot, OpenRouter, etc.). 3. In a session, reference files with
@, invoke Skills with/, or connect MCP servers (STDIO/HTTP/SSE). 4. Use the split-panel embedded terminal and browser to verify results. 5. Capture thoughts in the Ideas center, discuss them with the AI, and convert them into one-shot or Cron-recurring tasks tracked in the Task Center with status, run counts, and execution logs. 6. Optionally switch the runtime to Claude Code CLI / Codex CLI / Gemini CLI in Lab settings, or interact with the same work context via IM bots (Telegram/DingTalk). For development use./start_dev.sh; before contributing runnpm run typecheck,npm run lint,npm run test:classification, andnpm run test:unit.
What are this agent's strengths and limitations?
- Local-first architecture: sessions, tasks, config, and artifacts stay on your machine; Sidecars are isolated per Session and workspace file IO goes through Tauri/Rust rather than Sidecar HTTP
- Open on two axes: model providers are swappable, and the runtime can switch between the built-in Claude Agent SDK and Claude Code CLI / Codex CLI / Gemini CLI, avoiding vendor lock-in
- A complete task system: idea capture, Cron scheduling, local-command sensors (cheap condition checks that wake the AI only on match), execution logs, and failure recovery — all on one shared scheduler across Chat, IM bots, and background tasks
- Extensible tool ecosystem: MCP (STDIO/HTTP/SSE), built-in/user/workspace-level Skills, OpenClaw plugin bridge, and a
myagentsCLI - Embedded terminal and browser in a split panel close the execute-and-verify loop inside one window
- AGPL-3.0-only license: closed-source modification, OEM, or hosted offerings require a separate commercial license, adding compliance cost for business integration
- Heavy developer dependencies: self-building requires Node.js 22+, a Rust toolchain, and CMake 3.28+; Linux builds additionally need a set of WebKit/GTK system packages
- Model capability depends on external providers: you must configure your own API keys or subscriptions, and the external CLI runtimes are explicitly Lab-stage features
- The multi-process architecture (Tauri + Rust proxy + Node Sidecar + plugin bridge) is complex; contributors must read specs/tech_docs/ first, raising the customization bar
- Real provider/SDK smoke tests (test:credentialed) are excluded from default CI and must be run locally, so out-of-the-box reliability is partly on the user to verify
How does this agent compare with similar options?
The README positions the product as "not another chat window" and supports Claude Code CLI, OpenAI Codex CLI, and Google Gemini CLI as external runtimes — meaning it can act as a unified desktop shell for these command-line agents; the repo also tags "manus" as a topic, but no concrete comparison data is given.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MyAgents This agent | 57 · Major gaps | Desktop appFree + model costs | ★ 579 | 1d ago | TypeScript | Codex · Claude Code · Claude API |
| NextClaw | 58 · Major gaps | Desktop appFree + model costs | ★ 262 | today | TypeScript | Codex · Claude Code |
| HermesOffice | 57 · Major gaps | Desktop appFree + model costs | ★ 616 | 24d ago | TypeScript | OpenAI API · Claude API |
| OpenHuman | 51 · Major gaps | Desktop appFreemium | ★ 42k | today | Rust | — |
How does FollowAgents rate this agent?
Why each dimension lost points
README asserts local-first design, per-session Sidecar isolation, Tauri/Rust file IO, and Rust-proxied control plane under CORS/CSP; SECURITY.md advises reviewing tool permissions and stores keys locally in ~/.myagents/. However, no evidence shows a concrete permission model, tool-call confirmation flow, key encryption, or dependency auditing — these are assertions, so least_privilege/user_confirmation/sensitive_data_handling/dependency_security score 1. setup-no-egress.ts and CI notes show a strong test-layer zero-egress guard (externalResourceGuard, WidgetRenderer transport invariants), but that is a test constraint, not runtime protection; credentialed tests are properly separated from default CI. data_flow_transparency gets 2. No rollback/recovery mechanism is evidenced beyond task failure recovery mentions: rollback 1. AGPL plus third-party notices are complete: source_attribution 2.
README, package., and CI workflows are mutually consistent (version 0.4.15, Node 24, pinned rust-toolchain, matching layered test names): self_consistency 2. Lockfiles and --locked cargo commands support reproducibility: dependency_availability 2. No evidence in the provided files on error/failure message quality: failure_messages 1.
README explicitly names multiple audiences (developers, creators, researchers, PMs, educators) and concrete scenarios (workspace, task center, IM bots, cron): audience_and_scenarios 3. Capability boundaries are partially flagged (Lab runtimes, in-app model list authority) but thin: 2. Cron and low-cost command sensors with wake-only-on-match describe trigger precision adequately: 2. Explicit OS version requirements and per-platform setup notes: environment_fit 2.
Clear project structure and a full specs documentation system (architecture plus module docs): information_architecture 3. Detailed dev setup including platform differences, but end-user install relies on an external site: install_notes 2. Consistent naming (myagents, Mino workspace): 2. Screenshots and usage descriptions exist but no FAQ: examples_and_faq 2. Known limitations only scattered (Lab features): known_limitations 1. AGPL-3.0-only with LICENSING/COMMERCIAL/THIRD_PARTY_NOTICES and a verify:license script: license 3. No CHANGELOG evidence: versioning_changelog 1. Maintenance responsibility rests on one email and a personal repository; SECURITY.md timelines partially support it: maintenance_responsibility 1.
Feature tables and screenshots show usable output forms (task state, execution logs, audit): output_usability 2. Clear differentiation from plain chat windows: marginal_value 2. Cost/benefit claims are qualitative only, no data on API cost or resource usage: cost_benefit 1.
Architecture claims trace to named specs/ documents; CI test names map one-to-one to package. scripts: claim_traceability 2. Key facts (key storage path, test layering, version) cross-corroborate across README, SECURITY.md, package., and workflows: cross_source_corroboration 2. README distinguishes Lab features from stable ones and defers the model list to the in-app page, showing good fact/inference discipline: 2.
- This is a static review (low confidence); no builds or runs were executed, and test claims rest on CI configuration and script names only.
- Publisher is unverified and identity is unknown; the repo is maintained under a personal account with uncertain long-term commitment.
- No implementation evidence shows the runtime permission model, tool-call confirmation flow, or API key encryption; audit src-tauri and src/server security-boundary code before use.
- No CHANGELOG evidence; changes between versions cannot be traced from repository files.
- The release workflow is disabled in favor of local publish scripts; update-distribution auditability and signing chain should be verified independently.
- The xlsx dependency is pinned to a SheetJS CDN tarball rather than the npm registry; note the supply-chain provenance.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.5.5 v0.5.5 — 渠道访问控制与新旧会话可用性改进
Release notes excerpt
[0.5.5] - 2026-10-21 Added - 聊天机器人访问控制：支持白名单模式与自由模式，可按用户或群聊拉黑、解除拉黑；自由模式下自动登记实际使用的用户与群聊。 Changed - 渠道权限设置：统一白名单默认行为，重新整理用户绑定与群管理入口；微信扫码连接按平台实际能力展示设置。 - 自定义模型服务：创建服务商时可直接发现、添加和编辑模型并选择主模型，返回表单保留草稿，简化连接信息填写。 - 设置默认页面：从侧边栏或托盘打开设置时，默认进入模型服务商页面。 Fixed - 新对话可用性：渠道离线、旧会话正在执行或恢复时仍可独立新建对话，无需等待旧会话结束；旧任务与恢复继续进行，历史记录和渠道绑定保留。 - 会话切换与恢复：改善新旧会话切换时的消息显示和状态隔离，避免旧状态干扰新回复；修复进程重启或渠道换绑后使用旧连接信息导致的操作失败。 - 离线渠道管理：渠道停止后可正常重新启动，群权限可在离线时保存。 - 聊天滚动定位：修复历史定位与新滚动操作互相干扰、切换会话后旧定位仍生效，以及消息高度精度导致的底部偏差。 - 模型编辑弹层：修复弹层打开时的位置闪跳及输入框聚焦引起的页面滚动。View on GitHub ↗ -
v0.5.4 v0.5.4 — 聊天机器人跨聊天协作与会话稳定性改进
Release notes excerpt
[0.5.4] - 2026-10-18 Added - 侧边栏宽度调整：拖动侧边栏边缘即可调整宽度并自动记忆，聊天与工作区布局随可用空间适配。 - 聊天机器人发送模式：渠道可选择自动回复或仅通过工具发送消息，支持统一发送文字与附件。 - 跨聊天消息交接：Agent 可将消息、背景说明和附件交给同一 Agent 所连接的其他聊天，由目标会话结合已有上下文处理和转达。 Changed - 聊天机器人上下文：仅为实际绑定的聊天会话提供私聊或群聊、发言者、引用与历史消息等环境信息，普通桌面对话不受渠道发送指令影响。 - 渠道设置与名称展示：补全可获取的用户昵称和群名，按平台实际能力展示群消息接收选项，改善企业微信等插件渠道的平台识别。 - 自定义工作区记忆：初始化与后续记忆维护尊重工作区已有的记忆文件和组织方式，避免重复创建默认规则。 - CLI 信息展示：改善远程 Agent 长 ID 的可读性，准确区分渠道平台与插件名称，并明确历史查询范围。 Fixed - 会话与历史恢复：应用聊天界面意外重载后可恢复原有标签页；改善长历史会话重新打开或断线恢复时消息更新中断的问题。 - 聊天机器人响应与新建对话：修复部分会话启动或恢复时卡住，以及空闲一段时间后无法新建对话的问题。 …View on GitHub ↗ -
v0.5.3 v0.5.3 — 聊天机器人兼容性与完整历史恢复
Release notes excerpt
[0.5.3] - 2026-10-14 Added - 聊天机器人快捷入口：侧边栏「更多」新增「聊天机器人 Bot」，点击直接进入对应设置页；未开启协作空间时也可使用。 Changed - DeepSeek Harness 更新：内置运行环境升级至 0.1.18，支持删除历史无法读取的会话，改善故障恢复体验。 Fixed - 历史回答完整显示：修复工具调用较多时回答正文被截短的问题；长回答自动加载全文，重新打开会话与断线恢复时也保留完整正文。 - 微信插件兼容性：修复更新插件后因接口调整导致的启动与调用错误，保留旧版插件的兼容性。 - 企业微信二维码：修复二维码刷新后很快提示获取失败的问题，准确区分获取二维码与等待扫码时的错误。 - 设置页面跳转：从快捷入口进入设置或 Skills & Tools 时保留目标子页面，避免被重置到默认页。 - Goal 后台等待：等待后台工作完成时合理暂停续跑，避免无效重试与错误中断。 - MCP 工具恢复：改善连接失败后的重试与已停止工具的重新启动，准确展示工具状态；修复旧版配置导致的 Agent 或工具选择异常。 - 跨设备任务结果：短暂断线后可恢复结果回传，避免重复投递，并明确提示已不可用的会话。 …View on GitHub ↗