Hybrid-Computer-Use
llms.txt7 starsChanged 3 days agoMITView on GitHub
Hybrid Computer Use is an open-source MCP server, computer-use, that lets coding agents such as Codex, Claude Code, Gemini CLI and Grok Build operate Linux, macOS and Windows desktops. It adds two tools, AX (accessibility tree) and GUI (screenshots plus pointer and keyboard input), next to the agent's own shell and file tools, so the agent can mix code, accessibility and pixels in one task. Results reported in the paper (same model, harness, tasks and step budget; only the computer-use…
Files
# Hybrid Computer Use (HCU) > Hybrid Computer Use is an open-source MCP server, `computer-use`, that lets coding agents such as Codex, Claude Code, Gemini CLI and Grok Build operate Linux, macOS and Windows desktops. It adds two tools, `AX` (accessibility tree) and `GUI` (screenshots plus pointer and keyboard input), next to the agent's own shell and file tools, so the agent can mix code, accessibility and pixels in one task. Key facts: - Public MCP tools: exactly two, `AX` and `GUI`. Transport: stdio, newline-delimited JSON-RPC. - `AX` modes: `apps`, `observe`, `find`, `act`. The server names applications, windows and controls with short refs (`a1` -> `w1` -> `r12`). A ref stays bound to one window and fails as stale instead of rebinding. - `AX act` runs up to 64 ordered steps in one call, including `wait` steps (window, query, ref value or state, text, settle), and returns a compact diff of the accessibility tree. - `GUI` takes an ordered `actions` array: `screenshot`, `zoom`, `click`, `double_click`, `move`, `drag`, `scroll`, `type`, `keypress`, `wait`. Screenshots fit within 1366x768; coordinates are pixels of the returned image. - Guards: before input, `AX` checks that each ref still names the same live control, and `GUI` checks that the window under each coordinate still matches the last screenshot. A stale `GUI` batch fails with `gui_frame_stale`, sends nothing and returns a fresh screenshot. - One Go core with native providers: Linux (AT-SPI2, X11, XTest), macOS 14+ (AXUIElement, CoreGraphics, ScreenCaptureKit), Windows 10/11 (UI Automation via FlaUI UIA3, Win32). - Other entry points share one daemon per desktop: the `computer-use` CLI and the `hcu_sdk` Python package. - License: MIT; files adapted from OSWorld-V2 remain Apache-2.0. Results reported in the paper (same model, harness, tasks and step budget; only the computer-use interface changes; baseline CUA Driver): - OSWorld 2.0, 108 tasks, GPT-6 Astra: 72.37% Partial vs 68.82%. - OSWorld 1.0, 360 tasks, GPT-5.4: 70.13% vs 62.47%, 59.1% lower cost. - Windows Agent Arena, 111 tasks, GPT-5.4: 80.86% vs 71.89%, 50.6% lower cost. - MacAgentBench UI subset, 28 tasks, GPT-5.4: 25/28 vs 24/28, 51.4% lower cost. - WeaveBench, 114 tasks, GPT-5.4: 53.25% vs 46.37%, 46.3% lower cost. - OSWorld 1.0 across GPT-5.4, Claude Opus 5.5, Gemini 3.8 Flash and Grok 4.7: +6.48 points mean score, 17% fewer Steps, 30% lower cost. ## Docs - [Project page](https://microsoft.github.io/Hybrid-Computer-Use/) - [Paper (PDF)](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/paper/Hybrid-Computer-Use.pdf): "Hybrid Computer Use: A Unified Programmatic, Semantic, and Visual Interface for General Agents", 2026 - [README](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/README.md): overview, results, quick start, FAQ, citation - [Documentation guide](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/README.md): reading paths and terms - [Computer Use MCP](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/MCP.md): build, desktop runtime, agent configuration, `AX` and `GUI` contract, errors, troubleshooting - [Code entry points](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/CODE.md): daemon, CLI, Python SDK, programmatic tool calling, Codex code mode - [Native platforms](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/PLATFORMS.md): Linux, macOS and Windows providers and permissions - [Running OSWorld](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/OSWORLD.md): reproduce the OSWorld 1.0 and 2.0 evaluations - [Architecture](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/docs/ARCHITECTURE.md): Go packages, request path, AX and GUI state models ## Contract source - [tool_contract.go](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/computer-use/internal/contract/tool_contract.go): the published `AX` and `GUI` schemas and descriptions; `computer-use schema` prints them with the error table - [Agent instructions](https://github.com/microsoft/Hybrid-Computer-Use/tree/main/instructions): the base policy and the hybrid interaction guide given to agents ## Optional - [CHANGELOG](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/CHANGELOG.md) - [CONTRIBUTING](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/CONTRIBUTING.md) - [CITATION.cff](https://github.com/microsoft/Hybrid-Computer-Use/blob/main/CITATION.cff)
In this repositoryOne other file it gives its agents.
- AGENTS.mdAGENTS.md
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

