Headless background service
The AI is using the computer; no floating UI is ever shown.
Open source · Local first · Computer use runtime
Less waiting.
More doing.
FastCUA is a fast, reliable, and controllable local execution layer for computer-use agents. It brings interface state, native input, access policy, and human takeover into one resident control plane.
Runtime mind map
RUNTIMEFastCUA
M-01Observestructure + vision
M-02Executenative actions
M-03Controlpause + takeover
M-04ConnectMCP + host protocol
Use the right layer
for the right job.
FastCUA does not replace browser automation. It covers the Windows layer around it: applications, system dialogs, browser chrome, and workflows that cross app boundaries.
| Vision-first Computer Use | Browser bridge / extension | FastCUA | |
|---|---|---|---|
| Control scope | What appears in screenshots | Web pages inside the browser | Desktop applications and browser windows |
| Primary navigation | Pixels and coordinates | DOM · CDP · Browser APIs | UI Automation text; screenshots when needed |
| Model requirement | Usually needs vision | Usually text is enough | Text-only or vision-capable |
| Execution pattern | Often one action per loop | Browser commands | Multiple native actions per model turn |
| Human takeover | Varies | Usually limited to browser | Global pause, approval, interjection, and exit |
Speed comes from execution,
not marketing numbers.
FastCUA shortens the round-trip between an agent and a computer. This page describes mechanisms already present in the code—no invented product interface.
- P-01Observe01
Structured first, vision when needed
When accessibility information can express an interface accurately, read text and control structure directly. Request screenshots for canvases, custom controls, or visual verification.
- P-02Persist02
One warm control plane
Clients share a resident native host. Window identity, pointer state, approvals, pauses, and interruptions are not rebuilt for every individual action.
- P-03Execute03
Many actions per model turn
Stable clicks, keys, text input, scrolling, and drags can run in sequence. Observe again when layout, focus, or modal state changes.
Visible when it matters.
Interruptible at any time.
Full access is the default; unknown apps run without prompting. Set approvalPolicy back to safe in config.json to restore per-app approval.
Allow once, trust, or deny
Unknown applications cannot launch silently in safe mode.
Visible and reversible
Per-app prompts are disabled only while this mode remains on.
New actions are blocked
Resume in one step from the host control plane.
Host control plane
config.jsonSingle local config file
pause / resumePause / resume
interjectRedirect the current task
shutdownExit FastCUA completely
Pause, interjection, approval, and shutdown are exposed as named-pipe methods for host integrations. All settings live in the local config.json file.
A coherent control plane,
not a pile of clicks.
Window-aware coordinates
Actions stay attached to the target window and account for DPI scaling.
Text and pixels are independent
Request interface text, a screenshot, or both according to the next decision.
Native input
Clicks, key chords, Unicode text, scrolling, drag, and value replacement.
Two-way interruption
A person can pause or redirect; approval waiting also pauses the machine.
Exact trust rules
Canonical paths and executable names are matched exactly—never unsafe substrings.
Local first
MCP and the control plane use named pipes, configuration lives in a local file, and policy stays on the PC.
Platform capability stays in adapters.
Control semantics stay in the runtime.
A-01Clients
MCP-compatible Agents · Custom Clients
stdio MCPpersistent connection
A-02FastCUA Runtime
state · policy · approvals · interruption
control semanticshost protocol
A-03Host adapters
Native platform integration
current: Windows 11 x64The runtime defines consistent control semantics; host adapters provide windows, input, and platform APIs. Current release scope and future architecture boundaries remain separate.
What it is—and what
the current release is not.
- An open-source, local-first computer-use execution runtime
- An MCP tool layer for text-only and vision-capable models
- A control plane with pause, approval, and human takeover
- A complete cross-platform release package
- A replacement for DOM, CDP, or every specialized automation tool
- A way around UAC, secure desktops, or authentication surfaces
§ Real-world demoReal footage.
Real footage.
No concept mockups.
The Bilibili field-test video will show the complete path from task instruction and interface observation to execution and human takeover.
Video in preparation▶
BILIBILI TEST VIDEO
COMING SOONBring your own agent.
The installer prepares Node.js and the verified FastCUA runtime. It does not install, select, or modify an AI client.
irm https://raw.githubusercontent.com/Guojiz/FastCUA/main/install.ps1 | iex01 · Install runtime→02 · Open FastCUA Agent Setup.txt→03 · Give it to any MCP-compatible agent
The agent reads the bundled skill, configures its own stdio MCP connection, and verifies FastCUA. No specific agent is required.