基于快速代理构建的自托管人工智能助手，采用模型上下文协议（MCP）。组建多智能体团队，能够共同规划、研究、设计、编码、测试和部署。

状态： 正在开发中。公开发布谱系从v1.0.0开始。 核心架构详见AGENTS.md和 docs/SELF_HOSTING.md。

它的作用

• 多代理生成：一个由7个角色组成的敏捷团队（项目经理、业务分析、助理、开发、设计师、质量执行、数据管理官）协调代理间邮件和会议协议。
• 以MCP为先的工具：文件系统、GitHub、Atlassian、Figma、网页抓取（Scrapling）、Roborock vacuum、Google Calendar / Gmail、故事爬虫、TTS——全部作为MCP服务器公开。
• 免提语音：实时STT（更快的耳语、VAD、可选唤醒词）+ 流式TTS（Edge默认;ElevenLabs / OpenAI / Azure/系统选择加入）通过WebSocket，支持突入。故事锁定在免费Edge上，因此付费引擎不会消耗长文配额。
• Web 仪表盘：Vue + Vite UI，用于与 Jarvis 聊天、配置提供商、语音引擎、查看代理时间线、管理秘密。
• 自主机友好：单个docker compose up -d在Linux设备上调出整个栈。

看它实际行动

Jarvis

Self-hostable AI assistant built on fast-agent with the Model Context Protocol (MCP). Spawns a multi-agent team that can plan, research, design, code, test, and deploy together.

Status: active development. Public release lineage starts at v1.0.0. Core architecture is documented in AGENTS.md and docs/SELF_HOSTING.md.

What it does

• Multi-agent spawn: a 7-role agile team (PM, BA, SA, Dev, Designer, QE, DSO) coordinates over inter-agent email and meeting protocols.
• MCP-first tools: filesystem, GitHub, Atlassian, Figma, web scraping (Scrapling), Roborock vacuum, Google Calendar / Gmail, story crawler, TTS — all exposed as MCP servers.
• Hands-free voice: real-time STT (faster-whisper, VAD, optional wake word) + streaming TTS (Edge default; ElevenLabs / OpenAI / Azure / System opt-in) over WebSocket, with barge-in. Stories are locked to free Edge so paid engines never burn long-form quota.
• Web dashboard: Vue + Vite UI for chatting with Jarvis, configuring providers, voice engines, viewing agent timelines, managing secrets.
• Self-host friendly: single docker compose up -d brings up the whole stack on a Linux box.

See it in action

Not one chatbot — a whole team. Jarvis orchestrates a roster of specialist agents (research, finance, IoT, music, and more) — each one inspectable and directly reachable.

It remembers you — across conversations. Tell Jarvis a fact once; it saves to durable memory (not just chat history) and recalls it later in a brand-new conversation — no scrolling back, no re-explaining.

1. Teach it a fact:

2. Ask in a brand-new conversation — it recalls:

Chat with any specialist directly — not just the orchestrator. Switch straight to a subagent (here, the IoT agent) and ask your robot vacuum for its live status — real IoT (Roborock), not a mockup.

And it extends itself — live. No tool for the job? Jarvis scaffolds, installs, and wires in a brand-new MCP tool, then uses it in the same conversation — no package to hunt down, no redeploy.