This post compares five of them: Rakazo, OpenMausBot, Suparagent (formerly Open Duet), OpenTeam and kwant. Everything below comes from each project’s README or website as of November 13, 2026. Star counts move daily, so treat them as a snapshot, not a ranking.

Disclosure: I build kwant, one of the five. I’ve tried to describe the other four the way their own READMEs do, and to say plainly where they’re a better fit. If you spot something out of date, tell me and I’ll fix it.

What the hosted products actually are

It helps to be precise about what you’re replacing, because the three products differ more than the headlines suggest.

• Grok Bot (xAI, beta since August 11, 2026). You create named bots, message them like colleagues, and they hand work to each other in group chats. According to xAI’s documentation, as quoted in this explainer, all your bots share one persistent cloud computer per account, with one screen per bot, so logins and files are shared and a bot is not a security boundary. Access goes through eligible SuperGrok or Cursor plans, and there’s no self-hosted edition.
• OpenAI dots (launched September 29, 2026). An always-on agent inside ChatGPT that starts from your ChatGPT memory, uses Codex and connected apps, and works on its own cloud computer. OpenAI’s page lists Pro, Business Premium and Enterprise plans in eligible markets. At launch each user got one dot, with teams of dots described as a future plan (TechCrunch).
• Claude Duet (Anthropic, research preview since October 21, 2026). A proactive agent that sends a morning brief, watches your connected apps and asks before sending messages, paying, signing or deleting. It runs in a cloud VM with a browser. Access is a waitlist for US Pro and Max subscribers (The Verge, MacRumors).

All three keep the model, the computer and your data on the vendor’s side. That’s the trade the open-source projects undo.

What to compare

When I looked at these projects, five questions separated them more than feature lists did:

1. License. Can you run it for your company, change it, host it for others?
2. Where the bots work. A container desktop per bot, a shared browser, your own machine, or a cloud sandbox?
3. Which models. Only API keys, or also the Claude, Codex and Grok CLI subscriptions you already pay for? Local models?
4. Approvals. What does a bot do without asking, and what always stops for you?
5. How you run it. A desktop app, Docker Compose, a single binary, a phone app?

The comparison table

The projects, one by one

Rakazo

Rakazo calls itself an open-source, self-hosted alternative to Grok Bot, and also names OpenAI Dots and Claude Duet. Its feature list is the broadest of the five: routines, memory, computer and browser use, private or shared team computers, group chats, subagents, voice, and shareable dashboards and docs. The quick start pulls published images with Docker Compose, creates random secrets and starts the stack, with Postgres underneath. It’s the only one here with native iPhone and Android apps.

Pick it if you want the closest feature-for-feature copy of Grok Bot, phone apps included, and you’re happy running a Compose stack.

OpenMausBot

OpenMausBot (also MausBot, formerly OpenGrokBot) is the most-starred project in this space. Bots run on the claude, codex and grok CLIs already on your machine, so your existing subscriptions do the work. A local harness on 127.0.0.1 owns every agent process, and data lives in ~/.openmausbot. Each bot can use a cloud Linux desktop, an isolated local VM or, after explicit opt-in, your own computer, plus apps through Composio. It ships signed macOS builds, a Windows installer and an Ubuntu beta.

Pick it if you want a polished desktop chat app on Windows as well as Mac, built around the CLI subscriptions you already have.

Suparagent (formerly Open Duet)

Suparagent is a workspace for named AI coworkers with instructions, memory, routines and approvals. A lead agent can delegate to specialists, and the approval modes are the most granular here: Ask first, Accept edits, Approve for me, or Plan only. A routine never approves an action by itself. It supports OpenAI-compatible servers such as Ollama or LM Studio. The README is upfront that it’s early: the desktop app runs from a source checkout, and the self-hosted server is experimental and meant for one owner. The repository was created in 2023 and renamed, so its star count predates the current project.

Pick it if approval-first workflows matter most and you’re comfortable running from source.

OpenTeam

OpenTeam gives each bot its own Linux desktop with a browser, in Docker, on a server you run. Bots share a workspace and website sign-ins, keep memory, can be watched and taken over, and work in group chats or through subagents. The installer sets up a CLI and guided setup with private Tailscale HTTPS or public HTTPS. It recommends 8 GB of RAM. OpenTeam also published a hands-on test of six alternatives, which is worth reading alongside this post.

Pick it if you want a real desktop per bot on a home server or VM, and you’d like to watch bots work and step in.

kwant

kwant is a team of AI bots in one Rust binary, with SQLite and the React web app embedded. Bots share channels, hand work to each other with an @mention, and get it done on their own, on a schedule or while you’re away. Its distinctive pieces:

• “Done” means checked. Work shows up as cards on a Mission board, and a card turns Verified only when kwant runs the check itself, such as a test or a file check. A bot’s word or an attached file is never proof on its own. Since 0.9.0, a teammate on another model re-runs the checks in a clean copy of the files before you approve.
• Spaces. Separate spaces for work, a side project and home, each with its own bots, channels, routines, accounts and approval rule. A work space can use company API keys while a private space uses your own models.
• Every account you have. Several accounts per engine, with a priority order. When one runs out of usage, the turn moves to the next and the chat says so.
• Approvals with a hard floor. New spaces default to Ask only when risky: git push, network calls, unknown commands and changes outside the bot’s folder show the exact command with Allow once, Always allow or Deny. Payments always ask, in every mode.
• No telemetry. Messages go only to the AI providers and services you add.

It runs as an app on a Mac, on Linux and on Windows (both beta), or on a server as a single binary or with Docker Compose. There’s no native phone app, but the web app installs on a phone and sends Web Push notifications for approvals and questions. Release 0.9.5 shipped on November 13, 2026.

Pick it if you want bots whose results are checked before they’re called done, separate spaces for work and home, and something that runs on a laptop or a small VPS without a container stack.

Which one should you choose?

• Closest copy of Grok Bot with phone apps: Rakazo.
• Desktop app on Windows, Mac or Ubuntu with your CLI subscriptions: OpenMausBot.
• Approval before every change, a lead that delegates: Suparagent.
• A Linux desktop per bot on your own server: OpenTeam.
• Verified results, spaces, one binary on a laptop or VPS: kwant.
• You want the vendor to run everything: stay with Grok Bot, dots or Claude Duet. They’re the least setup, and that’s a fair reason.

Whichever you pick, the same caution applies: a self-hosted agent still sends your prompts to whichever model provider you configure. Self-hosting moves the computer and the data, not the model, unless you also run a local one.

Trying kwant

If kwant’s shape fits, the earlier posts in this series walk through it end to end:

• Self-Host a Team of AI Agents in One Binary covers your first team and the first @mention handoff.
• Install kwant on Linux: Ubuntu, Debian and AppImage covers the desktop app on Linux.
• Self-host AI agents on a VPS covers systemd, Docker Compose, HTTPS and safe defaults for an always-on team.

The quickest start is the Mac app, the Linux .deb or the Windows beta. Add an AI account you already have and ask your first question. kwant’s own comparison with Grok Bot goes deeper on that one matchup.

If you’d rather understand the machinery first, Tool Use in Go builds an agent loop by hand. Everything in this comparison is, underneath, that loop plus persistence, scheduling and approvals.

Next step: if you’re looking for an open-source alternative to Grok Bot, Claude Duet or dots, try one of these this weekend. If it’s kwant, star it on GitHub. It helps other people find it.