English | 简体中文
AUV means Application Use Via ....
- Apple Music Application Use Via
auv-apple-music... - macOS Media Control Use Via
auv-media-macos... - Balatro (yes the game Balatro) Application Use Via
auv-game-balatro... - ... more, waiting for your implementation.
Think of it as a programmable computer use, without agents.
- Getting Started
- Understand AUV
- Why even build AUV?
- Capability Matrix
- Development
- Related
- Acknowledgements
- Special Thanks
- Star History
- License
Install a prebuilt release:
brew install moeru-ai/tap/auv
auv --versionAlternatively, without Homebrew:
curl -fsSL https://auv.moeru.ai/install.sh | shcurl -fsSL https://auv.moeru.ai/install.sh | sh
auv --versionNote
Set AUV_VERSION or AUV_INSTALL_DIR to change the release version or the
install directory (default: ~/.local/bin).
The glibc Linux release is a full bundle: it includes AUV's private native
libraries, the PipeWire client files they need, and the English Tesseract
model. It needs glibc 2.36 or newer (for example Debian 12 or Ubuntu 24.04).
Additional OCR languages can be supplied through TESSDATA_PREFIX. Linux
musl is not currently supported.
scoop bucket add auv https://github.com/moeru-ai/auv
scoop install auv/auv
auv --versionDownload the archive for your architecture:
Extract the archive to a permanent directory. Add that directory to your user
PATH. The archive contains a single auv.exe; the Windows helper is embedded.
Install and configure proto first. Then add the AUV plugin and install the latest release:
proto plugin add auv "https://raw.githubusercontent.com/moeru-ai/auv/main/toolchain/proto/auv.toml" --to global
proto install auv latest --config-mode global --pin global
auv --versionNote
AUV Helper.app for macOS is included in the proto installation. On
Windows, auv-helper.exe is embedded in auv.exe and extracted only by the
elevated helper setup command.
Warning
Linux musl is not supported. (But PRs are welcomed!)
Install Nix 2.27 or later and enable the
nix-command and flakes experimental features. On macOS, install Apple's
build tools first:
xcode-select --installThen install the default AUV package from this repository:
nix profile install 'git+https://github.com/moeru-ai/auv#default'
auv --versionThe git+https transport is required so Nix fetches AUV's Git submodules. The
flake defines source-built packages for Apple Silicon and Intel macOS and for
x86-64 and ARM64 Linux. The package does not support Windows or Linux musl.
The Nix package does not embed the signed AUV Helper.app. On macOS, use
Homebrew, proto, or a direct release download if you need to run
auv setup macos-helper install with the official helper.
Prerequisites: Rust and the platform build tools below. AUV includes the Protobuf sources, so Buf is not required.
Warning
cargo install does not include AUV Helper. Use an install method above if
you need it.
Install the Xcode Command Line Tools:
xcode-select --install
cargo install --git https://github.com/moeru-ai/auv auv-cli --bin auv
auv --versionOn Ubuntu or Debian, install the native build dependencies:
sudo apt-get update
sudo apt-get install -y \
pkg-config libclang-dev libxcb1-dev libxrandr-dev libdbus-1-dev \
libpipewire-0.3-dev libwayland-dev libxkbcommon-dev libegl-dev \
libleptonica-dev libtesseract-dev
cargo install --git https://github.com/moeru-ai/auv auv-cli --bin auv
auv --versionNote
Other Linux distributions can use different package names.
Install Rust with the MSVC toolchain, Visual Studio Build Tools, and the Windows SDK.
cargo install --git https://github.com/moeru-ai/auv auv-cli --bin auv
auv --versionOfficial macOS releases include the signed AUV Helper.app. On macOS 13 or
later, install it for the current user:
auv setup macos-helper install
auv setup macos-helper statusTip
The installation does not require sudo or an administrator password.
If macOS requests approval, open the Background Items and Accessibility settings:
auv setup macos-helper open-background-items-settings
auv setup macos-helper open-accessibility-settingsNote
Projects that integrate AUV can rebrand AUV Helper.app. They can change its
name, icon, bundle identifier, and Apple Developer signing identity. See
Shipped helper identity
for packaging options.
Grant these permissions to the application that starts AUV, usually your terminal application:
| Permission | Needed for |
|---|---|
| Accessibility | AX tree reads, focused element control, keyboard/pointer automation. |
| Screen Recording | Screenshots, OCR, visual inspection, and evidence capture. |
| Automation | AppleScript/System Events app activation and foreground fallback paths. |
After you change the permissions, restart the terminal. Then run:
auv doctor
auv invoke permission.probeImportant
The Windows setup commands require an elevated PowerShell.
auv setup windows-helper install
auv setup windows-helper statusNote
The setup command extracts the auv-helper.exe that is embedded in auv.exe
into %ProgramFiles%\AUV. Then it registers that file as the LocalSystem
AuvHelper service. The Helper does not listen on the network. Lock and
unlock work through an ordinary auv serve that runs as the logged-in user.
To accept paired Devices from the network, start that daemon with
--remote-bind 0.0.0.0:9847. Setup also registers the AUV Daemon
scheduled task for the user who runs it. The task starts a local-only
auv serve at that user's logon, in the interactive session and with
limited privileges. If a 0.0.28 Helper is installed, run
auv setup windows-helper uninstall first, then install again and pair
clients again.
Evidence level: one installed lock and unlock gate on one Windows 11 host. This is not a general support claim. See the Windows Helper and daemon split.
auv invoke and auv mcp serve run commands through your local AUV daemon
and its Runner, so state such as a window maximize record stays between calls.
If no daemon runs, the first command starts one. It runs as you, accepts only
local connections, keeps its store and traces in your AUV state directory, and
stops 5 minutes after the last command. The command output names the Run and
the store. To run a command in the calling process instead, use
auv --no-daemon invoke <command-id>. AUV_ENDPOINT selects another daemon;
the CLI never starts one for it. On Windows over SSH (Session 0), the CLI
starts the AUV Daemon task in your interactive session instead, so run
auv setup windows-helper install first.
Use the instructions that match your installation method. If a platform Helper is installed, remove it first.
auv setup macos-helper uninstall
brew uninstall auvNote
Helper removal keeps the enrollment data in the login Keychain. It also keeps other AUV data in the Application Support directory.
rm "$HOME/.local/bin/auv"
rm -rf "$HOME/.local/bin/.auv"Run these commands from an elevated PowerShell:
auv setup windows-helper uninstall
scoop uninstall auvNote
Helper removal keeps the enrolled PINs in %ProgramData%. The daemon's own
pairing and policy data stays in its store.
If you installed the ZIP manually, remove its directory from the file system.
Then remove that directory from PATH.
cargo uninstall auv-cliFor Cua, agent-browser, and
similar computer-use projects, it is common to execute screenshot, read image, click, type,
wait, and follow-up verification steps in sequence, then ask LLMs or agents to judge the next move.
flowchart LR
A[Agent] --> B[screenshot]
B --> C[read image]
C --> D[decide next step]
D --> E[click]
E --> F[wait]
F --> G[type]
G --> H[verify]
H --> D
Many of those repeated sequences can be squashed into reusable GUI operations. Opening an app, waiting for readiness, filling a form, and checking the result should be callable as one command instead of spending tokens on the same step-by-step loop every time.
Modern agents often use skills or project instructions to orchestrate tool calls, CLIs, and scripts. But built-in computer-use surfaces, such as OpenAI Computer Use or Claude Computer Use, are still primarily interactive model-tool loops, not scriptable GUI automation libraries.
Similar to Playwright, what if we could organize those actions into executable scripts, reusable?
| Tool-call loop | Rust scripts |
|---|---|
|
pub fn open_and_fill_form(
app: &mut AppSession,
data: FormData,
) -> AuvResult<OperationResult> {
app.open()?;
app.wait_for_ready()?;
app.fill(data)?;
app.verify_submitted()
} |
|
pub fn scan_visible_rows(
region: &mut WindowRegion,
) -> AuvResult<ScrollScanArtifact> {
region.scan_rows_until_stop()
} |
|
pub fn verify_and_retry<F>(
mut operation: F,
) -> AuvResult<OperationResult>
where
F: FnMut() -> AuvResult<OperationResult>,
{
retry_until_verified(&mut operation)
} |
AUV expects agents to write, test, and improve reusable GUI automation for E2E tests and rapid application actions.
In fact, AUV is not a computer-use agent. It does not ship an agent or harness. It offers tools, CLIs, drivers, and verifiable observable results so agents can build reusable GUI operations.
AUV is meant to work with coding agents and agent products such as:
That means:
- If your agent can call a CLI, AUV can be used as computer use.
- If your agent can write code, AUV can move repeated GUI work into reusable Rust or JavaScript/TypeScript operations. Once a GUI flow is finalized as an operation, repeated execution can approach zero reasoning-token cost.
- AUV's daemon and extension APIs use versioned Protobuf/gRPC contracts. A language with compatible Protobuf/gRPC generators can generate a client for those contracts without AUV inventing another language-specific protocol. First-party SDK quality, packaging, and documentation are still separate support claims: Rust and JavaScript/TypeScript are available today, while a first-party Python SDK remains planned.
The reusable pieces are split by responsibility, but they use one execution model instead of becoming unrelated wrappers:
flowchart LR
A[CLI / MCP / Rust / JS / generated clients] --> B[typed operation]
B --> C[local or remote Device / Runner]
C --> D[capability Driver]
D --> E[direct result]
D --> F[Run trace and artifacts]
E --> G[separate semantic verification]
Drivers own platform capabilities, operation crates own reusable workflows,
and auv-tracing owns Run evidence and artifacts. The visual overlay remains a
separate trust and debugging surface; drawing a cursor never stands in for
input delivery or semantic verification. This package structure lets another
frontend or generated language client reuse the same operations rather than
reimplementing them around the CLI.
AUV born from the grounding knowledge of building general gaming agents for Project AIRI, since 2024, we tried to build agents to allow LLMs to play the following games, you can find how we implement the agents in the following repos:
There are more games we implemented where you can find in Project AIRI organization, but these four requires YOLO, OCR, screen understanding, and computer-use capabilities.
Now you have the framework to build for any applications, games.
Since Vercel published the agent-browser, we fell in love with it and have it assisted agents to build many web projects, but we found that the loop it requires for agents to call agent-browser CLI to execute the commands is too slow and inefficient, while in computer use world, many operations can be repeated thousands of times, just like how Playwright/Vitest would allow us to write E2E test for applications, why don't we expand this idea of writing code to control application to computer use world?
What AUV can do, compared to other computer-use projects.
- ✅: yes.
- ❌: no.
⚠️ : partial support. The cell states the limit.- ⏳: planned.
- —: not assessed.
Platform support comes from the Native desktop drivers row. Other rows name a platform only when their support is different.
| Capability | AUV | Cua | @oai/sky1bundled |
OpenBridge (KWWK core) | Playwright |
|---|---|---|---|---|---|
| Agent model | 💡 BYOA | 💡 BYOA | 💡 agent-free API | 💡 OpenBridge built-in agent KWWK is agent-free |
💡 BYOA + built-in Test Agents |
| Language-agnostic API | ✅ Protobuf/gRPC | ✅ HTTP/WebSocket | ❌ | ❌ | ❌ |
| Scriptable (Rust) | ✅ | ✅ | ❌ | ❌ | ❌ |
| Scriptable (TypeScript) | ✅ | ✅ | ✅ | ❌ | ✅ |
| Scriptable (Python) | ⏳ first-party SDK | ✅ | ❌ | ❌ | ✅ |
| Native desktop drivers | ✅ macOS/Linux/Windows ⏳ Android/iOS |
✅ macOS/Linux/Windows | ✅ macOS/Linux/Windows | ✅ macOS ❌ Linux/Windows |
❌ browser only |
| CLI | ✅ | ✅ | ❌ | ❌ | ✅ |
| MCP | ✅ | ✅ | ❌ | ❌ | ✅ browser MCP |
| REPL / Codemode | ⏳ planned | ❌ | ✅ Node REPL | ❌ | ❌ |
| Screen Lock/Unlock | ✅2 | ❌ | ❌ | ❌ | ❌ |
| Trace | ✅ Runs, artifacts, OpenTelemetry | ✅ trajectories | ❌ | ❌ | ✅ test traces |
| Screenshot | ✅ | ✅ | ✅ | ✅ | ✅ |
| OCR | ✅ macOS Vision/Linux Tesseract/Windows OCR | ❌ | ❌ | ❌ | |
| Template Matching | ❌ locator ✅ result contract |
❌ | ❌ | ❌ | ❌ |
| Accessibility tree | ✅ | ✅ | ✅ | ✅ | ✅ |
| Accessibility actions | ✅ | ✅ | ✅ | ✅ | |
| Mouse Click | ✅ | ✅ | ✅ | ✅ | ✅ |
| Mouse Move | ✅ | ✅ | ✅ Linux ❌ macOS/Windows |
— | ✅ |
| Background pointer input | ✅ macOS ❌ Linux/Windows |
✅ Linux window target ❌ macOS/Windows |
✅ | ✅ browser context | |
| Foreground pointer input | ✅ | ✅ | ✅ | ✅ | ✅ |
| Keyboard Hold | ✅ | ✅ | ✅ Linux timed hold ❌ macOS/Windows |
— | ✅ |
| Keyboard Input | ✅ | ✅ | ✅ | ✅ | ✅ |
| Scroll | ✅ | ✅ | ✅ | ✅ | ✅ |
| Ghost Cursor | ✅ macOS: multiple named cursors3 ❌ Linux/Windows |
❌ | ❌ | ❌ | |
| Customizable Cursor | ✅ macOS: colors, SVG, shadow ❌ Linux |
❌ | ❌ | ❌ | ❌ |
| Scroll-to-list | ✅ library and app integrations ❌ generic CLI |
❌ | ❌ | ❌ | ✅ browser lists ❌ desktop lists |
| Feedback | ✅ attempts, fallback, disturbance, verification | ✅ outputs and trajectories | |||
| YOLO / Custom Models | ✅ | ✅ | ❌ | ❌ | ❌ |
- Scroll scan is a major reason AUV exists. Most desktop automation stacks
can scroll and capture a screenshot. They do not make page records, row
candidates, crop artifacts, OCR fragments, or clear stop reasons. The current
scroll-scan implementation is contract work. The old
scan window-regionCLI will return when the reusable API is clear. - Feedback is machine-readable evidence for an action. It records the input path, changes, artifacts, fallbacks, and verification result. This evidence tells an operation when to retry, stop, or fail.
cargo fmt --check
cargo check
cargo testTo update vendored Protobuf dependencies, see the Protobuf source distribution reference.
Note
If you use proto, then
proto install buf
proto install node
proto install pnpm, this should help you install necessary tools.
pnpm install
pnpm generate:proto
pnpm exec playwright install chromium
pnpm build
pnpm test:run
pnpm lint
pnpm typecheckAfter you change headings in the root or package READMEs, run pnpm docs:update.
This command updates all README tables of contents.
Useful entrypoints:
auv doctor
auv invoke <command-id> --help
auv serve --help
auv devices list
auv runner --help
auv run --help
auv mcp serve
auv plugin listUse docs/TERMS_AND_CONCEPTS.md for shared vocabulary. Durable design and
evidence notes live under docs/ai/references/.
Note
This project is part of the Project AIRI ecosystem.
Special thanks to all contributors for their contributions to auv ❤️
Footnotes
-
Evidence level: installed package documentation and TypeScript declarations. The inspected package is
@oai/sky0.7.1 from the ChatGPT app. It is not available from the public npm registry. No native execution or native binary inspection supports this column. See the local Sky API research and the background-delivery comparison. ↩ -
Evidence level: configuration-specific installed-host test. This API locks and unlocks an existing login session. It does not sign in a user from the signed-out screen. After any OS restart, sign in to the Device once and start
auv serveagain; AUV does not install the daemon as a service yet. Then remote lock and unlock work for that session. The 2026-10-01 test ran 300 normal-use lock/unlock cycles. The API passed 298 cycles on the first attempt (99.33%). The requested OS state occurred on the first attempt in 299 cycles (99.67%). All 300 cycles ended in theUSABLEstate. The test used dwell times of 15, 20, 25, and 30 seconds. A separate stress test used delays near zero. It measured OS transition readiness, not normal-use reliability. Read the Device lock contract and platform evidence for the typed contract, native mechanisms, and configuration limits. The raw logs remain local. They are not in a durable evidence pack. ↩ -
AUV does not define a numeric cursor limit. Host memory and WindowServer resources limit the actual count. Ghost cursors are visual overlays. They do not deliver input or prove an action result. ↩