English | 中文
A multi-agent orchestration desktop application. Create agents, skills, and capabilities — wire them into collaborative workflows with 6 orchestration patterns.
- Direct Chat — Ask questions, get analysis and suggestions from the main assistant
- Team Orchestration — Dynamically compose multi-agent teams and run collaborative workflows
- Asset Creation — Create agents, capabilities, skills, and personas through natural conversation
- Long-term Memory — Remember preferences and context across sessions
- File Operations — Read, write, and search local files within a sandboxed directory
- Web Access — Real-time web search and page content reading
- Skill Packs — Import reusable knowledge packages (
.zip) with scripts and references - Canvas Editor — Visual workflow editor powered by ReactFlow for designing orchestration graphs
| Pattern | Description |
|---|---|
| Agent | Single agent execution |
| Sequential | Chain A → B → C |
| Concurrent | Parallel fork + aggregator |
| GroupChat | Round-robin or manager-selected speaker |
| Handoff | Relay-style task transfer |
| Magentic | Manager + workers |
- Shell: Electron
- Frontend: React, TypeScript, Vite
- Backend: Full TypeScript, Electron IPC (no HTTP server), SQLite (better-sqlite3)
- Canvas: ReactFlow
- Styling: CSS custom properties, glass-morphism design system
- LLM: Multi-provider support with thinking/retry/HITL
A self-built orchestration kernel — not a LangChain wrapper — engineered for semantic fidelity, resilience, and zero-trust security.
- Self-built Pregel engine — Wavefront/BFS superstep execution (emit at step N, deliver at N+1) with concurrent fan-out per layer, a 50-superstep infinite-loop fuse, and a fan-in barrier that survives partial node failures. Four orchestration patterns implemented with semantic fidelity; GroupChat ships all four correctness patches (cache, dedup, fairness, manager-output).
- Streaming intent router — A 24-char tail-buffer state machine decides in real time, across chunk boundaries, whether the main agent is answering directly or emitting a team-formation JSON command — so partial answers never leak to the user as text.
- Resilient LLM layer — The retry decorator wraps the actual request method (not the top-level call), making 429/5xx back-off impossible to bypass.
max_tokensis always sourced from agent configdefaultOptions(with an explicit 16384 default), never left to Anthropic's implicit API default — structurally eliminating its silent 1024-token truncation. Three-tier thinking support (policy gating → nativethinking_delta→<think>-tag fallback for proxies that tunnel reasoning as text, including orphan-close-tag recovery). - Multi-provider via one seam — A single
LLMProtocolinterface unifies Anthropic-native and OpenAI-compatible endpoints (DeepSeek, etc.) with shared retry, thinking, and tool-use semantics — including suspended-start tool-id reconciliation and cross-provider stop-reason normalization. - Three-tier memory — L0 identity → L1 rolling session summary → L2 cross-session digests → L3 on-demand fact retrieval. Each layer has a distinct injection point and injection discipline; L3 is a passive tool activated by prompt policy rather than hard-stuffed into context. The "no-vector on L3" decision is evidence-driven (measured FTS 59.1% > vector 36.4%).
- Zero-trust IPC — Every
ipcMain.handleroutes through awithHandlerwrapper returning a typedIpcResult<T>discriminated union; no uncaught rejection ever reaches the renderer. Atomic JSON writes (temp + rename) protect config, vault, and crash drafts from half-written state. - Crash recovery & self-healing storage — A
.runningsentinel + debounced draft persistence + a restore dialog closed the crash-recovery loop. SQLite runs WAL withintegrity_check, 30-min periodic backups, exit.bak, and corruption auto-restore (WAL checkpoint before copy — the subtle bit that prevents "fake backups"). - Sandboxed by default —
sandbox: true,contextIsolation: true,nodeIntegration: false; preload forced to CJS output (the sandbox has no ESM loader). LLM keys live in the OS keychain viasafeStorage— plaintext never round-trips to the renderer. - Hybrid vector knowledge base — FTS5 BM25 + flat-index vectors fused via RRF, with a transformers.js v4 WASM-in-Node recipe (ONNX runtime injection + custom cache + manual pooling) that runs cross-platform without native compilation.
- Orchestration-as-data — A content production pipeline ships as a serialized
Capabilitygraph, not code; users can open it on the canvas and rearrange the six-stage workflow. A GitHub-backed registry (no hosted backend) provides browse/import/auto-PR/one-click-update with non-destructive local-edit preservation and recursive dependency remapping. - i18n & theming from day one — multi-namespace locales with zh-CN/en kept fully in sync; main-process errors carry i18n keys for renderer-side translation. Brand color derivation runs in OKLCH perceptual space (not RGB shifting) to keep the glass aesthetic clean.
# Install dependencies
npm install
# Development
npm run dev
# Build
npm run build
# Package (DMG / installer)
npm run packagesrc/
├── main/ # Electron main process (IPC + tools + orchestrator, no HTTP backend)
│ ├── ipc/ # IPC handlers (agents, skills, capabilities, secrets...)
│ ├── llm/ # LLM client, retry, thinking parser
│ ├── orchestrator/ # Home router, workflow runner
│ │ └── patterns/ # 6 orchestration executors
│ ├── tools/ # Built-in tools (file, web, memory, create, askUser, opencli)
│ └── secrets/ # Vault (keychain-backed credential storage)
├── renderer/ # React frontend
│ ├── src/
│ │ ├── pages/ # HomePage, EditorPage
│ │ ├── components/ # MessageItem, MentionComposer, RunChatPanel...
│ │ ├── store/ # Zustand stores
│ │ └── api/ # IPC client, hooks
│ └── public/locales/ # i18n (zh-CN, en)
└── shared/ # Shared types (Agent, Skill, Capability, Persona...)
多 Agent 编排桌面应用。创建角色、技能与能力,通过 6 种编排模式组建协作工作流。
- 直接对话 — 主助手回答问题、提供分析和建议
- 团队编排 — 动态组建多 Agent 团队,运行协作工作流
- 资产创建 — 通过自然对话创建角色、能力、技能和人设
- 长期记忆 — 跨会话记住偏好和上下文
- 文件操作 — 在围栏目录内读写和搜索本地文件
- 联网搜索 — 实时全网搜索和网页内容读取
- 技能包 — 导入可复用知识包(
.zip),含脚本和参考资料 - 画布编辑器 — 基于 ReactFlow 的可视化工作流编辑器
| 模式 | 说明 |
|---|---|
| Agent | 单 Agent 执行 |
| Sequential | 顺序链 A → B → C |
| Concurrent | 并行分叉 + 聚合 |
| GroupChat | 群聊(轮询或管理员选发言人) |
| Handoff | 接力转交 |
| Magentic | Manager + Workers |
- 外壳:Electron
- 前端:React、TypeScript、Vite
- 后端:全 TypeScript、无 HTTP 服务层(Electron IPC 直连)、SQLite(better-sqlite3)
- 画布:ReactFlow
- 样式:CSS 自定义属性、玻璃拟态设计系统
- LLM:多供应商支持,含思考链/重试/人机交互
自研编排内核——非 LangChain 套壳——为语义保真、韧性与零信任安全而设计。
- 自研 Pregel 引擎——Wavefront/BFS 层式 superstep 执行(第 N 步 emit、第 N+1 步 deliver),同层并发扇出、50 superstep 防死循环熔断、容错 fan-in 栅栏。四模式语义保真落地;GroupChat 四补丁齐全(cache / dedup / fairness / manager-output)。
- 流式意图路由——24 字尾部缓冲状态机实时判定主助手是直答还是输出组队 JSON,跨 chunk 边界不漏判,直答片段永不误泄为正文。
- 韧性 LLM 层——重试装饰器包裹真正发请求的方法(非顶层调用),429/5xx 退避无法被绕过。
max_tokens始终取自 Agent 配置的defaultOptions(带显式 16384 缺省),绝不落空给 Anthropic 隐式 API 默认值——结构性杜绝 1024 token 静默截断。三层思考链支持(策略门控 → 原生thinking_delta→ 代理隧道<<think>>标签兜底,含孤儿闭标签恢复)。 - 单 seam 多供应商——一个
LLMProtocol接口统一 Anthropic 原生与 OpenAI 兼容端点(DeepSeek 等),共享重试/思考/工具调用语义,含 suspended-start tool-id 调停与跨供应商停止原因归一。 - 三级记忆——L0 身份 → L1 会话内滚动摘要 → L2 跨会话精炼 → L3 按需事实检索,各层注入点与注入纪律分离;L3 是被动工具靠 prompt 策略激活,不硬塞上下文。L3 不上向量为评测驱动决策(实测 FTS 59.1% > 向量 36.4%)。
- 零信任 IPC——所有
ipcMain.handle经withHandler包装返回判别联合IpcResult<T>,未捕获异常永不达渲染层。JSON 原子写(临时文件 + rename)保护配置、vault、崩溃草稿防半写态。 - 崩溃恢复与自愈存储——
.running哨兵 + debounce 草稿落盘 + 恢复对话框闭环崩溃恢复。SQLite 跑 WAL +integrity_check+ 30min 周期备份 + 退出.bak+ 损坏自动恢复(拷贝前先 WAL checkpoint——防"假备份"的关键细节)。 - 默认沙箱——
sandbox: true、contextIsolation: true、nodeIntegration: false;preload 强制 CJS 产出(沙箱无 ESM loader)。LLM 密钥经safeStorage存 OS keychain,明文永不同行至渲染层。 - 混合向量知识库——FTS5 BM25 + FlatIndex 向量经 RRF 融合,配 transformers.js v4 WASM-in-Node 配方(ONNX runtime 注入 + customCache + 手动 pooling),全平台通用免原生编译。
- 编排即数据——内容生产管线以序列化
Capability图交付而非代码,用户可在画布上重排六阶段工作流。GitHub 仓为后端的 Registry(无托管后端)提供浏览/导入/auto-PR/一键更新,本地修改非破坏性保留、依赖递归重映射。 - i18n 与主题从骨架做——多 namespace 语言包、zh-CN/en 保持同步;主进程错误带 i18n key 交渲染层翻译。品牌色派生走 OKLCH 感知色空间(非 RGB 位移),保玻璃态通透。
npm install
npm run dev