OpenMausBot Open-Sources the 'Grok Bot' Shape: An AI Team Inside One Chat Window
> Source: milind-soni/OpenMausBot — https://github.com/milind-soni/OpenMausBot > Language: TypeScript | Stars today: 1190 | License: Apache 2.0 > One-liner: Make AI assistants contacts in your address book — every bot gets its own personality, model, computer, and toolbox
One Assistant in One Box Is the Wrong Shape
OpenMausBot's author milind-soni argues that a single general-purpose assistant in a single window is the wrong shape for agents. Instead, the project borrows the structure of messaging apps: every contact in the sidebar is an independent AI agent with its own personality, model, computer, and connected apps. You talk to them like coworkers, watch them work, and approve actions at critical moments.
Context: Four Closed Products in Seven Weeks
In the first seven weeks of 2026, four closed personal-agent products shipped:
- Grok Bot (xAI): a roster of bots sharing one cloud computer, Grok models only, requires SuperGrok or Cursor subscription
- Muse (Meta): a single personal agent on a Meta cloud VM, US and Canada only
- dots (OpenAI): a resident agent inside ChatGPT, GPT models only, Pro or Business Premium required
- Cue (Manus): invite-only, agents with their own phone, email, and wallet
- Cloud Linux desktop: the bot's own cloud desktop
- Isolated local VM: a local virtual machine
- Your computer: on macOS and Ubuntu Xorg, after explicit opt-in, bots can drive your desktop (Ubuntu Wayland not yet supported, tracked in issue #345)
OpenMausBot is the open-source version of this shape: the same roster structure, but running on your machine, using your existing Claude, Codex, or Grok subscriptions, under Apache 2.0.
Three Key Design Decisions
1. Bring Your Own Agents
OpenMausBot provides no models. If claude, codex, or grok CLIs are installed, it uses them — existing logins and subscriptions work directly, no new accounts or proxying. You can run three bots with different models (e.g., Claude Opus for deep reasoning, Codex for codegen, Grok for fast search) and even switch models mid-conversation.
2. Local First
A small harness server runs on 127.0.0.1 and owns each agent process. Conversation history, keys, and events live in ~/.openmausbot — genuinely local, not "local-first with cloud sync."
3. Every Bot Has Hands
Each bot can use three compute environments:
Approval Cards: Supervision, Not Autonomy
When a bot wants to run a shell command, edit a file, or ask a sensitive question, an inline card appears in chat: Allow / Deny / Answer. This runs counter to the "autonomous agent" narrative — a permission broker turns every risky operation into a decision you make, equally for cloud and local machines. OpenMausBot gives you a highly capable intern, not a fully automated employee: strong, but every critical action needs your sign-off. That friction is precisely what makes it safe to let agents touch your production environment.
Jev: A Fast Decision Model
If a Jev key or OpenRouter key is configured as the Decision Model, a room can be set to Auto mode: when an untargeted message arrives, the lightweight decision model picks the right bot within a few hundred milliseconds, falling back to the room's lead bot on mistakes. Not every message needs the most expensive model — a routing layer sends heavy work to the right specialist.
Teams as Files
OpenMausBot can install an entire team from a single Markdown file. Templates on BotMRR can be added with one click; the app opens a review screen showing the bots, channels, playbooks, connector checklist, and suggested routines to be created. YAML frontmatter powers reliable one-click installs, while the plain Markdown playbook stays readable by Grok, Claude, ChatGPT, and humans. Connections default to off, routines default to paused, and packages carry no credentials, conversations, permissions, memories, or computer access.
How It Differs from CopilotKit/OpenBot
| Dimension | OpenMausBot | CopilotKit/OpenBot | |------|-------------|---------------------| | Form | Desktop chat app | Web platform | | Deployment | Local .dmg/.deb/.exe | Docker Compose | | Models | Bring your own CLI (Claude/Codex/Grok) | Bring your own API key | | License | Apache 2.0 | AG-UI protocol | | Governance | Approval cards | Gateway auditing | | Positioning | Personal agent team | Enterprise AI coworkers |
OpenMausBot is closer to "WeChat for AI" — a chat app on your machine where every contact is an independent agent. CopilotKit/OpenBot is closer to "Jira for AI" — an enterprise-deployed work platform where every coworker is a governed agent. Same direction, different entry points.
Why 1,190 Stars Today
OpenMausBot's surge is less about technical breakthroughs and more about timing. After a wave of closed agent products in early 2026, developer sentiment crystallized: I don't want my agent team locked inside some vendor's cloud. OpenMausBot doesn't offer a more powerful agent — it offers a shape: AI as contacts in your address book, running on your own machine, on subscriptions you already pay for. Four major companies validated the demand; OpenMausBot simply open-sourced it.
---
Project: https://github.com/milind-soni/OpenMausBot Website: openmausbot.com License: Apache 2.0