Every Contact in the Sidebar Is an Agent on Your Machine. The Interesting Part Is the Permission Card.

OpenMausBot added about 7,000 stars this week, which puts it in the top ten of GitHub's weekly chart next to repos with a hundred times its age. It is an open-source desktop chat app where each contact in the sidebar is a real agent, running on the CLIs already installed on your computer, with its own model, its own personality, its own computer and its own connected apps.
The pitch is that four closed products shipped this shape in seven weeks of 2026, and every one of them keeps the model, the computer and your data on the vendor's side of the line. OpenMausBot is the Apache 2.0 version of that shape.
That framing is honest about what it is: a reimplementation of a category, not an invention. The question worth answering is whether the reimplementation gets the parts right that the closed products get to hide.
Two processes and one loop
The architecture is small enough to describe in a paragraph, which is a compliment.
A React frontend renders the chat, the model picker and the computer panel. It holds no transports of its own; it sends typed commands over HTTP and folds one server-sent event stream into state. Behind it, a harness server on `127.0.0.1:8799` owns every agent process, normalises each provider's native protocol into one canonical event stream, and logs it per thread as NDJSON. Transcripts, keys and events live in `~/.openmausbot`.
Agents are the `claude`, `codex` and `grok` CLIs you already have. It uses your existing logins and subscriptions, with no proxy in the middle and no new account to create. You can point an engine at a custom CLI binary in settings if you run a pinned build.
That choice has a consequence worth naming. Because the agents are the vendor CLIs, OpenMausBot inherits both their capability and their terms. It is a shell around someone else's harness, and shells around someone else's harness break when the harness changes.
Three kinds of hands
Each bot can act through one of three computers: a cloud Linux desktop, an isolated local VM, or your own machine.
The cloud desktop spins up on demand, shows a live screen preview while the bot works, and can be taken over in a browser. The local VM keeps the work on your hardware. Host control, meaning the bot operating your actual computer, is available on macOS and Ubuntu Xorg only after explicit opt-in.
Ubuntu Wayland is explicitly disabled for host control, with the reason stated in the README: a real-seat safety gate is unresolved and tracked as issue #345. Legacy opt-ins get cleared. Chat, preview, cloud and local VM all still work on Wayland.
That paragraph is the most credible thing in the repository. Shipping a feature and disabling it on the platform where the safety model does not hold, in public, with an issue number, is how engineering works and almost never how marketing does.
Connected apps come through Composio, with a one-click marketplace over Gmail, Slack, GitHub, Notion, Linear and hundreds more. OAuth once, and every bot can use them as tools.
The permission broker is the product
Shell commands, file edits and questions surface as inline cards: allow, deny, or answer in chat. The README calls it a permission broker, and it applies to cloud and local computers alike.

This is where the design earns its keep. An agent with a cloud desktop, a Gmail token and a shell is a general-purpose actor with your authority. The only thing standing between a plausible-sounding instruction and a deleted mailbox is a decision surface that makes the risky step visible and cheap to refuse.
Most agent products treat approval as a modal to be dismissed. The detail that suggests someone thought about it here is that questions are also cards. When a bot needs a value it does not have, it asks in the conversation rather than guessing and proceeding.
There is a second, less comfortable detail. Application API keys are stored as plaintext, not encrypted, in a private `config.json` with 0600 permissions on Unix. The README says so directly instead of implying otherwise. For a personal tool on a laptop, that is a defensible trade. For a shared machine, it is a reason to use the local VM instead of the host.
What is still missing
The status section is unusually candid. The loop works end to end: message, agent, streamed reply, tools, approvals, computer use. macOS, Windows and Ubuntu 24.04 x64 have released builds, with Ubuntu still labelled beta and carrying the capability limits above.
Hosted and mobile connectivity is still being built. Webhook triggers currently use a local receiver on `127.0.0.1:8800` that exposes only a health check and secret hook paths, which means the app has to be running to accept a delivery. Public delivery requires proxying that receiver through a relay or a tunnel. Phone access is optional, requires pairing, and cannot use a localhost link.
Voice is split the same way. Hosted voices need an ElevenLabs, Fish Audio or xAI key. The built-in Mac voices and the local Chatterbox model need no cloud key. On Windows and Linux, speech-to-text routes through Fish Audio.
Each of those is a real constraint and each is stated. The aggregate effect is that the product is honest about being early, which is more useful to a prospective user than a feature grid.
Why this shape is spreading
A personal agent needs four things: a model, a computer, a set of credentials and a place to keep context. Every product in this category is an answer to "where do those four live".
Closed products put all four on the vendor's side and sell convenience. That works until the agent needs to touch a file that cannot leave, or a session that must not be logged, or a tool that has no API. Then the boundary becomes the product's limitation rather than its feature.
The local-first answer is that all four live on your machine, and the cost is that you now own an operations problem: four subscriptions, three CLIs, a permission model, and a config file with plaintext keys in it.
OpenMausBot's contribution is showing what that operations problem looks like when it is written down. The harness on localhost, the normalised event stream, the approval cards, the platform-specific safety gates: those are the boring decisions that determine whether a personal agent is a toy or a tool you leave running.
It will be copied. The architecture is too obvious once you see it, and Apache 2.0 makes copying legal. What will be harder to copy is the habit of shipping the caveats alongside the feature, and that habit is the only reason to trust the thing with a shell.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.