Markus AI Workforce
Turn a plain-language goal into coordinated work by a self-hosted team of specialist AI agents.
- Source repo
- markus-global/markus
- Stars
- ★ 219
- Last updated
- today
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 51/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Researchers who need competitor scans, market sizing, or due diligence with sources cited.
- Not a fit if
- Production teams requiring a stable 1.0 release
- Users who only need a single chat assistant
- Website
- markus.global ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.12.4 (2026-11-14)
- Source review
- 51/100 · Major gaps
What does this agent do, and when should you use it?
Markus is an open-source AI workforce platform that users run themselves: describe a goal, and it assembles a team, divides the work, and reports back. Its components include a TypeScript agent runtime and context engine, REST and WebSocket APIs, a React dashboard, an Electron desktop app, and SQLite storage. Tasks can run in parallel, dependent work waits for prerequisites, and deliverables go through peer review. Users can work through the desktop app or run the CLI on a server and open its UI in a browser. Markus is pre-1.0 at version 0.11.x, and the README warns that minor releases may include breaking changes.
After a user submits a goal, Markus creates a requirement, assembles a role-based team, and breaks the work into tasks with dependencies. Agents work in separate workspaces and review one another's deliverables; once approved, blocked follow-up work can start automatically, and results appear on the board with notifications. The repository also includes GUI and browser automation, communication bridges for Slack, Feishu, WhatsApp, Telegram, and Discord, plus an agent-to-agent protocol. The README lists model options from Anthropic, OpenAI, Google, DeepSeek, MiniMax, Fireworks, OpenRouter, and Ollama, and skill imports from skills.sh, SkillHub, OpenClaw, AgentScope, or any MCP server.
- Researchers who need competitor scans, market sizing, or due diligence with sources cited.
- Content teams turning one brief into an article, social thread, newsletter, and short-video script.
- Engineering teams assigning parallel development work to agents, reviewing deliveries, and merging as humans.
- Operations staff who want recurring price, risk, or inbox checks to report only when something changes.
How do you install or deploy this agent?
The desktop app is the recommended entry point; download an installer for macOS, Windows, or Linux from the project website or GitHub Releases. For a headless machine, install the CLI on Linux or macOS and start it; alternatively, install it with npm in a Node.js 22 or later environment. SQLite and the UI ship with the app, so no separate database installation is needed.
How do you use this agent?
Start the desktop app or open the browser UI, then give the Secretary a concrete goal in plain language, such as market research and a launch plan. Watch the agents work, steer them, and approve their results. For server use:
What are this agent's strengths and limitations?
- Turns a goal into dependency-linked tasks that can run in parallel.
- Uses peer review and approval gates instead of relying on the executing agent to report itself done.
- Can run on your own machine, keeping control of runtime location, data, and keys; SQLite and the UI are bundled.
- The README names multiple model providers, including local models through Ollama.
- The project is pre-1.0 at 0.11.x, and minor releases may contain breaking changes.
- The CLI path requires Node.js 22 or later; model providers may require their own services or credentials, but setup details are not included in the supplied material.
- The README does not specify model usage costs, permission scopes, or agent failure recovery behavior.
- Communication bridges and MCP skill imports are described, but their configuration steps are not provided.
How does this agent compare with similar options?
The README contrasts Markus with a single copilot: a copilot handles one task at a time, while Markus coordinates specialist roles in parallel, retains memory, patrols a task board, and applies peer review.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Markus AI Workforce This agent | 51 · Major gaps | Desktop appFree + model costs | ★ 219 | today | TypeScript | OpenAI API · Claude API |
| Switch AI Team Workspace | 70 · Some gaps | Desktop app | ★ 4.7k | today | Python | — |
| DoraBot Desktop Assistant | 45 · Major gaps | Desktop appFree + model costs | ★ 248 | 4d ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| Bitterbot Desktop | 55 · Major gaps | CLIFree + model costs | ★ 2.5k | 1d ago | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence describes multi-agent orchestration, browser automation, messaging integrations, and coding tools. The Claude Code and Codex skills explicitly describe non-interactive execution that automatically approves edits and commands, while the README promotes unattended monitoring; least privilege, confirmation, and external-effect controls therefore score 1. The configuration document separates secrets from ordinary settings, documents precedence, and describes 0600 secret files; the security policy says the API binds to loopback and routes deny by default, supporting a score of 2 for sensitive-data handling, though this sample does not show the enforcement implementation. The packet does not sufficiently explain where task data goes, and gives little evidence on dependency security or rollback, so those score 1. A private reporting channel is documented, but publisher identity is unverified and the skill's originalSource is only a template field, so attribution scores 1.
Lazy CLI loading, centralized communication exports, and the single-source configuration document show some consistency. However, the configuration record marks fixes as in progress, the A2A export is deprecated, and a test documents that automatic delegation cannot select an agent when no skills are requested; self-consistency scores 1. CI specifies Node 22 and pnpm 9 and includes several test gates, but coverage steps are allowed to fail and the sample cannot establish availability of external model, browser, or messaging services, so dependency availability scores 1. Examples include failure causes and retry guidance, and the security policy gives response targets, but runtime failure handling is only partially evidenced; failure messages score 1.
The README targets users who want an AI team and lists research, content, software, and monitoring scenarios, with desktop, CLI, VPS, cross-platform, and multi-model options. Audience and environment fit score 2. The skills document Chrome version and connection-mode limitations, and the coding-tool skills describe automated permission modes and suitable tasks, giving capability boundaries and trigger precision 2 each; framework-wide permission setup and mode selection still depend on the operator.
The README links to architecture, API, user guide, engineering records, and contribution guidance, and includes quick installation, examples, a folded server option, and pre-1.0 caveats. Information architecture, installation notes, and examples score 2 because some claims are not implementation-verified by this packet. The package is versioned 0.12.3-rc.0 and the README warns of breaking changes before 1.0; CI and dated architecture records provide update context, but no single changelog is evidenced, so versioning scores 2. Naming stability scores 1 because package descriptions vary. The full Apache-2.0 text is included and the README explains dual licensing and skill-specific licenses, earning 3. Contribution and security policies establish maintenance and reporting routes, but the responsible publisher identity is unverified, so maintenance responsibility scores 2.
The product presents task decomposition, parallel work, peer review, persistent memory, and a delivery board as a coherent workflow, and the README gives a readable delivery example. Output usability and marginal value score 2. No model-call budget, resource use, or cost data is provided, so cost-benefit scores 1.
The README links feature claims to architecture and guides, and the configuration document names paths, precedence, and implementation locations, giving claim traceability a 2. The packet includes CI, tests, and source, but they do not independently substantiate the README's broad effectiveness claims, so cross-source corroboration scores 1. Documents distinguish target, implemented, in-progress, and residual work, and the test record discloses a failure and limitations; however, the packet also contains self-reported benchmark claims that are not independently verified, so fact/inference separation scores 2.
- The Claude Code and Codex skills describe automatic approval of file edits and commands without interactive review. Enable them only for trusted tasks and isolated workspaces, and review the permissions granted to external tools and browser automation.
- The README's around-the-clock monitoring, compounding memory, and self-development benchmark are not independently verified by this static packet. The product is pre-1.0, and the configuration record marks changes as in progress.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Desktop App (recommended) | Platform | Download | |----------|----------| | macOS (Apple Silicon) | [Markus.dmg](https://github.com/markus-global/markus/releases/download/v0.12.4/Markus-0.12.4-arm64.dmg) | | macOS (Intel) | [Markus.dmg](https://github.com/markus-global/markus/releases/download/v0.12.4/Markus-0.12.4.dmg) | | Windows x64 | [Markus-Setup.exe](https://github.com/markus-global/markus/releases/download/v0.12.4/Markus-Setup-0.12.4.exe) | | Linux x64 | [Markus.AppImage](https://github.com/markus-global/markus/releases/download/v0.12.4/Markus-0.12.4.AppImage) | Server / CLI …View on GitHub ↗ -
Release notes excerpt
Desktop App (recommended) | Platform | Download | |----------|----------| | macOS (Apple Silicon) | [Markus.dmg](https://github.com/markus-global/markus/releases/download/v0.12.3/Markus-0.12.3-arm64.dmg) | | macOS (Intel) | [Markus.dmg](https://github.com/markus-global/markus/releases/download/v0.12.3/Markus-0.12.3.dmg) | | Windows x64 | [Markus-Setup.exe](https://github.com/markus-global/markus/releases/download/v0.12.3/Markus-Setup-0.12.3.exe) | | Linux x64 | [Markus.AppImage](https://github.com/markus-global/markus/releases/download/v0.12.3/Markus-0.12.3.AppImage) | Server / CLI …View on GitHub ↗
FAQ
Do I have to buy Markus?
Can I run it on my own machine?
Can I use a local model?
Is Markus ready for deployments that require a stable release?
Is Markus AI Workforce free?
Is Markus AI Workforce open source?
Is Markus AI Workforce safe to use?
Is Markus AI Workforce still maintained?
Who makes Markus AI Workforce?
What is Markus AI Workforce's official website?
Thanks — we use this to improve the page.