Announcements
Braus 1.0-4: one local memory for your agents
Braus 1.0-4 brings local app knowledge and interaction memory into a shared archive for workers, MCP clients and the CLI. Here is what the release includes.
Braus 1.0-4 is available for download. It gives agents a desktop workspace on your Mac: create a worker in the app, or connect Claude Code or Codex through local MCP to work in desktop software.
This release brings the app's Brain, interaction memory and procedure recall into the release baseline. Workers and local clients can draw on shared app knowledge while keeping their own task streams. The agent still observes the current interface and checks what changed after acting.
The direction remains the one in our GitHub README: let agents work in your apps while you keep using your Mac. Here is what the release contains and how to start a small trial.
A workspace for tasks inside Mac apps
Braus's Driver moves app windows to a virtual display within the current Mac session and routes the agent's input there. The workflow uses your existing apps and files.
You can create several workers. They take turns controlling desktop apps, with the others waiting for access. The virtual display shares the current user session and filesystem; it does not create a separate virtual machine.
Support remains specific to the app and workflow. The README lists AppKit, Chromium, Electron and Qt, with UXP support for Adobe apps still in development. Start with the operation you need and check its output on your installed configuration.
From the current window to structured context
Perception turns the interface into structured text, using pixels and accessibility information when available. That gives the agent context for finding a target even when an app exposes little useful structure through accessibility.
Braus refreshes that view as the window changes. The engine uses the live interface to resolve an action and check its effect. A previously known control can still move, disappear or become ambiguous in a new state.
This makes observation and verification part of the workflow. A worker should be able to show the result, including where an effect remains uncertain.
Brain and Memory now share a local foundation
Brain maps app controls, relationships and transitions from observations and action evidence. Memory records interactions, outcomes and learned procedures in a local SQLite archive.
That archive is shared by Braus workers, local MCP clients and the CLI on the same Mac. A client's task stream remains private to its grant. Shared app knowledge and an individual client's current task have different scopes.
For a later task, Recall can suggest parts of an earlier procedure, together with conditions and evidence. The agent must resolve those steps against the current interface and verify their results. Recall remains in development and can be inconsistent; we are not announcing a measured reliability or token-saving gain from memory in this release.
| Component | What it contributes | What still needs checking |
|---|---|---|
| Brain | Context about controls, relationships and transitions | The current window and target |
| Memory | Recorded interactions, outcomes and procedures | Whether the earlier evidence fits the task |
| Recall | Suggested steps from relevant earlier work | Conditions, parameters and observed results |
The current memory documentation explains this local sharing. It does not provide cross-device or team synchronization.
Bring your existing agent through MCP
The app includes the local bridge. Braus and the external client run on the same Mac, and Braus stays open while the connection is in use.
Open MCP → MCP Connections…, create a client entry, enable it and wait for Ready. Copy the generated Claude Code command, Codex command or JSON from Client configuration. Use those generated values for your installation.
Each connection entry accepts one active client. Create a separate entry for independently running clients. Start with the observation-only check in the local MCP guide.
For an existing setup, regenerate the configuration from the installed Braus app rather than carrying over a helper path from another installation. Keep earlier app data intact and follow the current documentation for compatibility checks.
New benchmark results alongside the release
We have also published a new driver comparison dated 13 October. Across ten paired app reads, Braus returned 67.5–95.4% fewer estimated full-read tokens than Cua Driver 0.34.0. The median was 351 versus 2,156, approximately 84% lower.
The report uses a measured backend candidate, separate from the released installer. It includes operation checks and five tasks driven by a real model. Braus completed three of those task attempts; Cua direct completed five. Compact response context and completed tasks remain separate outcomes.
The benchmark report provides the configurations, measurements and limits. The source revisions make it possible to distinguish a published result from the app version you install.
Start with a read-only task
The v1.0-4 release page contains the Braus DMG. The project targets macOS 15+ on Apple silicon and Intel; workflow compatibility still needs checking on your app, hardware and OS build.
Install and open Braus, connect a supported model and grant the Mac access required for desktop work. The README suggests reading an open TextEdit test document and summarizing it without editing or saving.
Confirm that the agent identified the intended document and read its contents correctly. Then try a bounded edit in disposable data with a finish condition you can inspect. Keep the app version, starting state and any manual intervention with the result.
Braus is free and open source under Apache 2.0. Model access has its own requirements. Local window-to-text processing can feed a remote model, so review the permissions and data guide for the connection you use.
Download Braus 1.0-4, choose one small task and inspect the output. Bring a reproducible example to GitHub Issues when the app reaches a boundary. Those examples are the work that makes the next release better.