Markus

An open-source AI workforce you run yourself.
Give it a goal in plain language — it builds the team, splits the work, and reports back.

English | 中文

For people who want an AI team, not a chatbot.

Watch it work

You     I need a competitive analysis and a go-to-market plan.
Markus  ▸ requirement created
        ▸ team assembled — 3 roles, 3 tasks, 2 dependencies
Research Lead       competitor scan     reviewer: Senior Researcher
            Senior Researcher   market sizing       reviewer: Research Lead
            Content Director    go-to-market plan   blocked by the two above
08:41  two tasks start in parallel, each in its own workspace
        09:12  competitor scan delivered → peer review → one revision requested
        09:26  revision delivered → approved
        09:30  go-to-market plan unblocked, starts automatically
        09:58  three deliverables on the board, you get one notification

One copilot is an intern. You need a company.

A copilot is brilliant at one task, remembers nothing tomorrow, and grades its own homework. The work that actually matters — a quarter of research, a product launch, a codebase that grows — is not one task.

| | A single copilot | Markus | | -------------------- | --------------------------- | ----------------------------------------------- | | Scale | One task at a time | Specialists running in parallel | | Memory | Gone when the session ends | Persists, consolidates, and compounds | | Initiative | Waits for your prompt | Patrols its own task board, around the clock | | Quality | "Done" is self-reported | A peer reviews and gates every delivery | | Visibility | N tabs, N windows | One board, one audit trail | | Where it runs | Someone else's cloud | Your machine, your keys, your data |

Models and skills: any provider — Anthropic, OpenAI, Google, DeepSeek, MiniMax, Fireworks, OpenRouter, or a local one via Ollama — and skills in both directions: import from skills.sh, SkillHub, OpenClaw, AgentScope or any MCP server, export your own back.

🚀 Quickstart

The desktop app is the recommended way in. It brings its own runtime and its own browser, so there is no Node.js to install, no terminal, and nothing to open afterwards.

| Platform | Installer | | --- | --- | | macOS — Apple Silicon or Intel | Markus-….dmg | | Windows — x64 | Markus-Setup-….exe | | Linux — x64 | Markus-….AppImage (also .deb, .tar.gz) |

Download from markus.global or GitHub Releases — the same files either way.

Then open the app and give your Secretary something real to do:

"We're launching in Europe next quarter. Research the market, size it, and draft a launch plan."

You watch, steer, and approve.

# Linux / macOS — installs a runtime too if Node.js is missing
curl -fsSL https://markus.global/install.sh | bash && markus start
Anywhere with Node.js 22+
npm install -g @markus-global/cli && markus start

No database to install, no cloud account to create: SQLite and the UI ship with the app.

_{Working from source instead? See CONTRIBUTING.md.}

What people put it to work on

• Research and analysis — competitor scans, market sizing, due diligence, with sources cited.
• Content operations — one brief becomes an article, a thread, a newsletter and a short-video script.
• Software — this repository is built this way (see below).
• Standing watch — daily scans, price and risk monitoring, inbox triage. The kind of work that is mostly "check, then report" runs unattended and wakes you only when something actually changed.

It builds itself

Markus is developed on Markus. The issues, requirements, task assignments, peer reviews and release notes in this repository run through a Markus organization — the same product you download. Bugs its own agents find get fixed by its own agents and merged by a human.

It is the most honest benchmark we have: if it could not ship itself, you should not trust it to ship your work.

Under the hood

A TypeScript monorepo: the agent runtime and context engine, a REST + WebSocket API, a React dashboard, an Electron desktop app, SQLite storage, comms bridges (Slack, Feishu, WhatsApp, Telegram, Discord), GUI and browser automation, and an agent-to-agent protocol.

The internals are documented properly — start at the documentation index:

| | | | --- | --- | | Architecture | How the pieces fit, and why | | Agent runtime · Memory · Tools | The three subsystems people ask about most | | API reference · User guide | Build against it, or just use it | | Engineering records | Dated audits and post-mortems, published unfiltered |

💬 Community

• GitHub Discussions — questions, show & tell, case studies:
• Blog — tutorials and product notes:
• Discord — real-time help with users and contributors — coming soon
• 微信群 — 中文用户交流群，内测与贡献支持（建设中）

Contributing

pnpm install && pnpm build
pnpm dev          # API + web UI in dev mode
pnpm test         # unit + integration tests
pnpm typecheck    # TypeScript, all packages
pnpm lint         # ESLint

• Good first issues — small, well-scoped, mentored
• Help wanted — things the community needs
• Bug reports — with a repro, ideally

Markus is pre-1.0 (0.11.x) and moving quickly, so minor releases can contain breaking changes. That is precisely why feedback, issues and pull requests are most valuable right now.

License

Markus is dual-licensed:

• Open source — Apache-2.0. Use it, modify it, self-host it, ship it commercially.
• Commercial — available for teams that need support, indemnification, OEM embedding or custom terms.

Website · Blog · Discussions · Issues

_{Markus — where AI agents work as a team}