An open-source desktop app where AI agents work on your local projects,
and every file they read, command they run and line they change stays in view.
Download for Mac · Documentation · 简体中文
Point wuu at a folder, describe the task, and let an agent read the code, run commands and edit files. The conversation shows each step as it happens. Beside it, Review shows the actual diff on disk, so you check the result rather than taking the summary on trust.
In the 18-task Terminal-Bench 4 subset reported in PR #630, Wuu matched Codex's pass rate at 11.0% lower recorded API cost, and solved one more task than Pi at 6.6% lower cost.
| Harness | Passed | Pass rate | Recorded API cost (USD) | Cost per success (USD) |
|---|---|---|---|---|
| Wuu | 10/18 | 55.6% | $32.40 | $3.24 |
| Codex (Code Mode) | 10/18 | 55.6% | $36.42 | $3.64 |
| Pi | 9/18 | 50.0% | $34.68 | $3.85 |
All three harnesses used GPT-6 Astra with high reasoning effort: 18 tasks each, 54 runs total, with skills, memory, ask-user and subagents disabled. Codex 0.150.1 and Pi 0.84.4 retained their native prompts. Codex enabled Code Mode (code_mode=true, code_mode_only=true); Wuu exposed direct tools alongside optional Code Mode composition. Costs include cached input and exclude title generation; cost per success is total recorded API cost divided by passed tasks.
Connect OpenAI, Anthropic, DeepSeek, xAI, Kimi, Z.AI, MiniMax or Alibaba Cloud Model Studio with an API key. Gateways such as OpenRouter and SiliconFlow and local servers such as LM Studio work too, as does any custom OpenAI- or Anthropic-compatible endpoint. You can also reuse a subscription you already have: a Codex login, xAI SuperGrok or Grok Build.
Already use another coding agent? wuu can run Claude Code, Codex, Cursor, Devin, DeepSeek Harness, Grok, Hermes, OpenCode, Pi or Antigravity as the conversation's engine, with that agent's own login and tools. Settings → Subscriptions shows plan quotas and balances for every account in one place.
The workspace panel sits beside the conversation:
- Review shows the current Git diff, file by file.
- Files previews code, Markdown, images and documents.
- Terminal opens a shell in the project.
- Browser is the same page the agent drives. A floating preview follows its clicks, and clicking or typing on the page yourself pauses the task.
Select text in a file to add it to the conversation, leave a comment, ask about it in a side chat, or have the agent edit just that passage. Responses work the same way: quote a passage into your next message.
Computer Use lets the agent observe and operate native macOS apps through Accessibility and native input. A floating preview follows its pointer and shows each click and keystroke while the target app stays in the background. The agent saves through the app itself, so Review shows the real file change. Screen Recording and Accessibility permissions apply.
Press Enter while a task runs to steer it, or Command + Enter to queue the next message. Fork from any earlier message into the same folder or a new Git worktree. Start a conversation in its own worktree so experiments stay off your branch. Use /side for a quick question that stays out of the main thread, and search to find any past conversation.
Pick a permission mode for each conversation. Read-only is for investigation. Standard allows edits inside the workspace. Unconfined lifts the boundary when you really need it. In Read-only and Standard, the agent's commands run inside the macOS Seatbelt sandbox. Turn on Approve for me to have the model review shell commands and risky actions before they run.
When the work belongs somewhere else, run the agent's tools in Docker, over SSH, in Singularity/Apptainer, or in a Modal, Daytona or Vercel sandbox.
Plugins add tools and features: scheduled automation, persistent goals, task checklists, memory, subagents and more. Each one can be turned on or off. You can install a local plugin, add skills, connect MCP servers and run hooks. The official TypeScript SDK gives external and bundled plugins the same public API. See Extend wuu.
The desktop preview runs on Apple silicon Macs. Download it from GitHub Releases, move wuu.app to /Applications, and open it.
The release pipeline requires Apple Developer ID signing and notarization. Older previews used ad-hoc signatures; check the release notes and installation guide for the downloaded version.
On first launch, connect a model service and add a project folder as a workspace. Start with a small task, then check the changes and test results in Review. The quick start walks through an example.
The desktop app includes its own core. To use wuu from a terminal or script as well, install the Go version listed in go.mod and build the CLI:
git clone https://github.com/blueberrycongee/wuu.git
cd wuu
make install
wuu initMake sure Go's binary directory is on your PATH, and configure a model provider. Then run a task:
cd /path/to/your/project
wuu exec --permission-mode read_only "review this project and explain how to run its tests"The wuu exec guide covers scripting, JSONL output and session controls.
wuu reads and changes local files and runs commands within the active permission mode. Prompts and relevant context go to the model provider you choose. Sessions and settings are stored under ~/.wuu by default. Read the security model before working with sensitive data or untrusted projects.
See Contributing to work on wuu, or open an issue to report a problem. Report security vulnerabilities through SECURITY.md.
Licensed under MIT. Agent avatars use blobatar, also MIT-licensed.
Wuu's macOS Computer Use implementation draws on the design of Cua Driver, an MIT-licensed project by Cua AI, Inc. Thanks to its contributors for their work on native computer use.