There’s a specific kind of clutter that comes with running AI agents locally: the terminal window problem. One tmux pane streaming a Claude session you set up Tuesday, another running Codex, a third that’s just scrollback nobody will ever read again. You don’t have a team of agents — you have a pile of terminals with attention deficit. OpenMausBot takes a swing at this with a simple reframing: your agents should be contacts in a chat sidebar, each with a name, a job and their own working environment.
What it actually is
OpenMausBot is an open-source Electron desktop app — TypeScript-strict, React 19 — for macOS, Windows and Ubuntu. Under the hood every “bot” in the sidebar is a real Claude or Codex agent running locally. You bring your own models and subscriptions; the app gives each agent a face, a chat thread, and slots for connected apps. It positions itself as the open-source answer to Grok Bot, with Meta Muse, OpenAI dots and Cue by Manus named as the commercial products it’s chasing.
The distribution polish is better than most hobby desktop projects: signed and notarized macOS builds for Apple silicon and Intel, a Windows x64 installer, and a Ubuntu 24.04 .deb with an AppImage beta. Source is on GitHub under a permissive posture that welcomes PRs.
The contact list is doing real work
The obvious criticism is that this is cosmetics — a chat skin over a CLI. That misses the part that matters. The killer feature isn’t that agents have avatars; it’s that each contact gets its own everything: its own personality, its own model, its own cloud computer, its own connected apps. That’s not a skin, that’s isolation.
Anyone who’s run a single omnibus assistant knows the failure mode: the context where you asked it to summarise invoices bleeds into the session where it’s refactoring your dotfiles, and by Friday it’s confidently wrong about both. Splitting agents into named contacts with separate environments makes delegation look like what it actually is — different collaborators, different contexts, different permissions. A research bot that reads the web doesn’t need the same hands as a code-review bot that never leaves the repo. Making that split literal, at the app level, is the right call.
The other idea worth keeping is approval as a first-class interaction. OpenMausBot treats “approve what matters” as a designed moment rather than a config flag. If you’ve ever watched an agent confidently do the wrong thing at speed, you know the difference between a tool you trust and a tool you babysit is exactly one confirmation gate, placed well.
The homelab angle
- Local-first by design. The agents run on the models you already have. There’s no hosted middleman between you and your own agent fleet — consistent with how the rest of a self-hosted stack should look.
- It pairs with a dedicated agent machine. Run it on the homelab desktop or an old laptop and your “team” lives where your other services live, not in some SaaS tab you’ll lose the login for.
- It’s a good chassis to learn agent ergonomics on. If you’re still figuring out how many agents is too many, the contact-list model makes the cost of each new one visible: every contact is another persona, environment and permission set to maintain.
The honest cautions
An Electron app wrapping agents that have real access to your machine is a thick trust sandwich. Before you connect a bot to anything that matters, read what it can actually touch — the same discipline you’d apply before handing a new hire keys. And treat each contact’s connected apps like firewall rules: grant the minimum, per agent, not one broad grant to the app. The project is also young and independent (it was renamed from OpenGrokBot), so expect churn; pin a version that works and update deliberately, the way you would with any early infrastructure tool.
Verdict
OpenMausBot is an opinionated bet that persistent, named, isolated agents beat one generic assistant and a dozen dying terminals. For people who already run local Claude or Codex and want the interaction model to catch up to the capability, it’s a clean, single-installer way to make that bet. Even if you never install it, steal the two ideas: per-agent environments, and approval gates that are part of the design, not an afterthought. That’s the shape of competent agent work in 2026 — the terminal scrollback was never going to survive it.

