An open-source AI team that ships while you sleep.
You give it a goal in plain language. It hires the team, splits the work, runs
everyone in parallel, reviews every delivery, and keeps going while you rest.
And yes — it built itself.
English | 中文
🪞 Dogfooded into existence. The Markus project is built on Markus: issues, tasks, code, reviews, releases — our own agent team runs the entire loop on itself, start to finish. If it can ship itself, it can ship whatever you're building.
- Not a wrapper — agents talk to LLM APIs directly and use real tools: shell, files, git, web search, code analysis, GUI & browser automation, any MCP server.
- Ships 24/7 — a heartbeat keeps the team moving, reviewing, and escalating. You sleep; they ship.
- Memory that compounds — three-layer persistent memory, auto-consolidated between sessions. The team gets measurably smarter the longer it runs.
- Your data, your machine — fully self-hosted. SQLite by default (PostgreSQL supported), zero mandatory cloud, zero lock-in.
A single copilot is a smart intern: great at one task, forgets everything overnight, and calls its own work "done." One employee doesn't make a company.
| Single copilot | Markus team | |
|---|---|---|
| Scale | One task at a time | Parallel work across specialist roles |
| Memory | Evaporates when the session ends | Persistent, auto-consolidated |
| Initiative | Waits for your prompt | Heartbeat patrols tasks 24/7 |
| Quality | "Done" is self-reported | Peers review and gate every delivery |
| Visibility | N tabs, N windows | One dashboard, one audit trail |
# Desktop app (macOS / Windows / Linux)
# → https://github.com/markus-global/markus/releases/latest
npm install -g @markus-global/cli # Node.js 22+, or the Linux one-liner without Node
markus startOpen http://localhost:8056 — the onboarding wizard creates your account (initial login: admin@markus.local / markus123). Then tell your Secretary:
"I need a research team: scan our competitors, write a competitive analysis, and draft a go-to-market strategy."
Markus assembles the team, breaks the goal into tasks, and starts executing — specialists in parallel, every delivery reviewed.
That's it. SQLite + bundled web UI, zero external dependencies. From source: git clone → pnpm install && pnpm build && pnpm dev.
- 🧠 Three-layer memory — procedural, semantic, episodic. Knowledge accumulates across sessions and consolidates on its own.
- ⏰ Heartbeat-driven initiative — open tasks, async completions, and blockers keep moving even with no one watching.
- 🔀 True concurrency — multiple sessions run in parallel on isolated per-session workspaces. No cross-talk, even inside one chat.
- 🧬 ContextOS context engine — pinned structural anchors, a stable context budget, and compression that never drops decisions. Long, busy sessions stay fast and grounded.
- 🛡️ Trust & gates — progressive trust levels, a formal submit → review → merge lifecycle, full audit trail, emergency pause.
- 🔌 Skill ecosystem — import skills from skills.sh / Claude Code, SkillHub, OpenClaw, AgentScope, and MCP servers — and export your best ones back.
- 🤖 Any LLM — Anthropic, OpenAI, Google, DeepSeek, MiniMax, Ollama, OpenRouter, and more — with unified model discovery and automatic failover.
- 🔒 Bring your own keys — credentials live in your deployment, never in a third-party cloud.
Full skill details: Skill Ecosystem
┌─────────────────────────────────────────────────────────┐
│ Web UI (React) · Desktop (Electron) │
│ Dashboard · Chat · Projects · Builder · Hub │
└──────────────────────┬──────────────────────────────────┘
│ REST + WebSocket
┌──────────────────────┴──────────────────────────────────┐
│ Org Manager (API Server) │
│ Auth · Tasks · Governance · Projects · Reports │
└──────────────────────┬──────────────────────────────────┘
│
┌──────────────────────┴──────────────────────────────────┐
│ Agent Runtime (Core) │
│ Agent · LLM Router · ContextOS · Tools · Skills · │
│ Memory · A2A · Concurrency · Decision · Heartbeat │
└──────────┬────────────────────────────┬─────────────────┘
│ │
┌──────────┴──────────┐ ┌────────────┴─────────────────┐
│ Storage (SQLite / │ │ Comms (Slack, Feishu, │
│ PostgreSQL) │ │ WhatsApp, Telegram) │
└─────────────────────┘ └──────────────────────────────┘
TypeScript monorepo with modular packages:
| Package | Role |
|---|---|
| core | Agent runtime — LLM routing, ContextOS, tools, skills, memory, concurrency, heartbeat, workspace isolation |
| org-manager | REST API, WebSocket, governance, task lifecycle |
| web-ui | React + Vite + Tailwind dashboard |
| desktop | Electron desktop app (macOS / Windows / Linux) |
| cli | @markus-global/cli — one-command install and launch |
| storage | SQLite persistence (zero external dependencies) |
| gui | GUI automation — VNC, screenshots, input control, visual analysis |
| comms | Slack / Feishu / WhatsApp / Telegram bridges |
| a2a | Agent-to-Agent communication protocol |
| remote | Remote access — tunnels and zero-config networking |
| chrome-extension | Browser automation via the Markus extension |
| shared | Shared types, constants, utilities |
| Guide | Description |
|---|---|
| User Guide | Setup, configuration, Web UI walkthrough |
| Architecture | System design, agent runtime, memory, governance |
| Agent Runtime | Agent lifecycle, execution model, workspace isolation |
| Tool System | Built-in tools, MCP integration, tool contracts |
| Skill Ecosystem | Import/export skills from skills.sh, SkillHub, OpenClaw, AgentScope, MCP |
| Memory System | Three-layer memory architecture (Tulving) |
| Cognitive Architecture | Cognitive Preparation Pipeline (CPP) design |
| Mailbox System | Agent attention model, priority queue, triage |
| Prompt Engineering | System prompt assembly, tool loop, compression |
| State Machines | Task & requirement FSM specification |
| Concurrent Processing | How one agent handles multiple sessions / mailbox items in parallel |
| Streaming & Reattach | Streaming events, reconnection, tool-loop integrity |
| API Reference | REST API endpoints and WebSocket events |
| Coding Tools | Claude Code / Codex / Cursor integration |
| Learning Loop | Agent self-improvement and memory consolidation |
| Remote Access | Cloudflare Tunnel, Tailscale, FRP, ngrok setup |
| Release & Distribution | Build, packaging, publishing pipeline |
| Blog | Articles and tutorials on Markus and AI agents |
- GitHub Discussions — questions, show & tell, case studies: https://github.com/markus-global/markus/discussions
- Blog — tutorials and product updates: https://markus.global/blog
- Discord — real-time chat with users and contributors (English/global) — coming soon
- 微信群 — 中文用户交流群，获取帮助、内测与贡献支持（建设中）
Join details and the contributor escalation path are in docs/COMMUNITY.md. All channels follow our Code of Conduct.
pnpm install && pnpm build
pnpm dev # API + Web UI in dev mode
pnpm test # Run tests
pnpm typecheck # TypeScript check
pnpm lint # ESLint- Good first issues — beginner-friendly tasks
- Help wanted — features the community needs
- Bug reports — help us fix issues
See CONTRIBUTING.md for full guidelines.
Markus is dual-licensed:
- Open Source: Apache-2.0 — free to use, modify, distribute, and self-host for any purpose, including commercial use
- Commercial: Available — for teams needing enterprise support, indemnification, OEM embedding, or custom terms
Skills shared through the marketplace may use their own licenses (typically MIT).
Website · Blog · Discussions · Issues
Markus — Where AI Agents Work as a Team