Claude Code can edit your repo and run your tests, but it can’t get into the DaVinci Resolve timeline you’re cutting. Codex can write the script, but you’re still the one clicking through the export dialog, zipping files in Finder and sending them on Slack. A huge amount of work on a computer has a UI but no API, and coding agents still can’t reach it.
Braus wants to fill that gap. On macOS it gives the agent its own “desk”: the target window is moved to a virtual display and operated there, while your mouse and main screen stay free for other things. You can create workers inside Braus, or connect Claude Code or Codex to it through a local MCP server. Details checked on October 17, 2026.
What problem it solves
The official demo shows two complete workflows (edited and sped up; the real runs take on the order of 25 minutes):
- In DaVinci Resolve, turn a clip black and white, export it, zip it in Finder and send it on Slack.
- In Chrome, search X for posts about computer use and compile them into a draft.
What they have in common: they span several desktop apps, dialogs and export flows, with no single API in between. Braus’s pitch isn’t “one more chat box”; it’s turning the software you already have into tools for an agent.
What the product is
- Repo: ForteAI-Org/braus (about 21 stars at the time of writing, per the GitHub API; it’s new, created on September 9, 2026)
- License: Apache-2.0 (including the Driver, Perception and Engine libraries, which you can embed in your own macOS projects)
- Site: braus.ai · Announcement: Introducing Braus · Docs: braus.ai/docs
- Latest build: release v1.0-4 (October 13, 2026),
Braus-1.0-4.dmg(the older v1.0-1 assets still used the previous name, Mecum) - Platform: macOS 15+. It installs on Apple silicon and Intel, but the team says Intel hasn’t been validated; the documented test machine is an M4 MacBook Pro on a specific OS build
- Implementation: Swift 6.4, with topics including computer-use, mcp and desktop-automation
The core breaks down into roughly four parts:
| Module | What it does |
|---|---|
| Driver | Moves app windows to a virtual display and routes the agent’s input there, trying not to steal your main cursor |
| Perception | Turns the current UI into structured text (pixels plus whatever accessibility info is available), using fewer tokens than feeding raw screenshots to a model |
| Brain | Builds up a map of each app’s controls, relationships and navigation as you use it |
| Memory | Stores interactions and results in a local SQLite database; Recall is still being polished, and the team admits it isn’t consistent yet |
Current framework support: AppKit, Chromium, Electron and Qt, with Adobe UXP in progress. For Catalyst and OpenGL-style interfaces, the official benchmarks show very low success rates on effect checks, so don’t expect it to control everything out of the box.
Installing it and hooking up MCP
- Download the DMG from Releases, install it and open Braus.
- In the app, choose Create Worker, connect a model provider, then Allow Mac Access (Screen Recording, Accessibility and so on; approve each system prompt).
- For your first run, use a throwaway test document (the team suggests TextEdit), and spell out which app, what to change, where to stop and what not to touch.
- To use Claude Code or Codex: keep Braus open, go to MCP → MCP Connections…, add a client, wait until Enabled shows Ready, and copy the command the app generates (don’t hand-copy an outdated template).
For Claude Code it looks roughly like this (the path and UUID come from what the app generates):
claude mcp add --scope user --transport stdio braus -- \
"/Applications/Braus.app/Contents/Helpers/braus-bridge" \
mcp-bridge --connection \
"/Users/YOUR_USERNAME/Library/Application Support/Braus/MCP/CLIENT_UUID.connection.json"
Codex uses the same bridge via codex mcp add braus -- …. The client and Braus must be on the same Mac; the bridge speaks STDIO, so there’s no Node or Python to install. Each connection serves one active client, and you shouldn’t share your *.connection.json files.
Model costs are separate: Braus itself is free, and the model side can be a subscription, an API key or a local option like Ollama. Turning windows into text happens locally, so remote models receive structured text and task context (some tools may still send screenshots, depending on how the connection is configured).
A few details worth remembering
- Multiple workers: you can create several, but control of the desktop is taken in turns. It’s not truly parallel clicking.
- Focus stealing: some apps still grab focus when they launch or pop up a dialog. Braus tries to restore your window, and if that fails the run may pause.
- Stop doesn’t undo: stopping a worker doesn’t roll back anything it already clicked. Keep an eye on side effects like exports and sends yourself.
- Token comparison: compared with Cua Driver 0.34.0’s full-screen “screenshot plus accessibility tree” reads, the team measured roughly 67.5% to 95.4% fewer estimated tokens (a median of about 351 versus 2,156). That’s context size per tool response, not your bill or your task success rate.
- Seat status: unvalidated setups show Seat: Not Validated. Actions are still recorded; just be aware of it.
How it differs from screenshot-based computer use
Many tools send a full-screen screenshot to a multimodal model and click on coordinates. Braus uses a virtual-display workstation plus structured UI text: the agent’s cursor doesn’t take over your main screen, and tool responses are claimed to be leaner. The catch is that Perception has to understand each specific UI framework, which explains why the official compatibility table is so restrained instead of claiming “works with every Mac app”. For your first task, deliberately pick a read-only summary in TextEdit, check the tool list with /mcp, and only then move on to tasks with side effects like exporting.
What’s good
- It genuinely targets computer use for desktop software without APIs, rather than just wrapping a terminal.
- Open source down to the Driver layer under Apache-2.0, so it’s easy to audit and integrate.
- The local MCP setup generates its config in one click, so it fits into existing Claude Code or Codex workflows with little disruption.
- Perception uses structured text, which saves context compared with raw screenshots (under the official benchmark setup).
What’s lacking
- Still early in stars and ecosystem (about 21 stars when I checked). Few issues also means a thin community of people who’ve hit the same problems.
- Hard requirement of macOS 15+; Intel and most OS version combinations aren’t validated.
- Perception misses controls, and compatibility depends on the app, the version and the task. Success rates on complex UIs (game engines, some GPU-rendered interfaces) aren’t pretty.
- Long tasks (like editing and exporting video) take a while, pauses mid-run need someone watching, and Stop doesn’t undo.
- Recall in Brain and Memory is “still inconsistent”, so don’t treat it as a reliable SOP engine.
Who it’s for, and who it isn’t
Good fit: Mac users already on Claude Code or Codex whose day keeps getting stuck on export dialogs, design software or clients without an API, and who are willing to verify in small steps with test documents and can live with early-software rough edges.
Not a fit: people whose main machine runs Windows or Linux; anyone who only wants to chat in a browser and never grants desktop permissions; and anyone expecting one sentence to reliably drive any commercial suite. Read the compatibility notes before installing.
If you only have agents edit code and run tests, stick with the terminal. If video editing, office clients or internal Electron tools eat half your day in manual clicking, Braus is worth an afternoon: start with a read-only TextEdit task to get MCP working end to end.
Links
- Repo: https://github.com/ForteAI-Org/braus
- Website: https://braus.ai/
- Announcement: https://braus.ai/blog/introducing-braus
- Docs: https://braus.ai/docs
- Releases: https://github.com/ForteAI-Org/braus/releases
