@injaneity/pi-computer-use

pi-computer-use

Codex-style computer use for Pi on macOS.

pi-computer-use gives Pi agents a semantic computer-use surface for visible macOS windows. It prefers Accessibility (AX) targets such as @e1, returns semantic state after every action, and attaches screenshots only when AX coverage is too weak for reliable operation.

Table of Contents

• Quick Start
• What It Adds to Pi
• Examples
• How It Works
• Documentation
• Development & Benchmarks
• Release & Install Notes
• License
• See Also

Quick Start

Install the Pi package:

pi install git:github.com/injaneity/pi-computer-use@v0.2.5

Start Pi in interactive mode. On the first session, grant macOS permissions to:

~/.pi/agent/helpers/pi-computer-use/bridge

Required permissions:

• Accessibility
• Screen Recording

Some browser automation paths use JavaScript from Apple Events. If the browser blocks that, Pi surfaces a model-readable hint asking the user to enable Allow JavaScript from Apple Events in the browser's developer menu, then retry.

Then call screenshot first in a Pi session. It selects the controlled window and returns the latest semantic state, including AX refs such as @e1 when available. If the target app/window is ambiguous, use list_apps and list_windows first.

list_apps()
list_windows({ app: "Safari" })
screenshot({ window: "@w1" })
click({ window: "@w1", ref: "@e1" })
set_text({ ref: "@e2", text: "hello" })

Use /computer-use in Pi to inspect the effective config and config sources.

What It Adds to Pi

• Public tools: list_apps, list_windows, screenshot, click, double_click, move_mouse, drag, scroll, keypress, type_text, set_text, wait, arrange_window, navigate_browser, computer_actions.
• AX target refs in tool results, with capabilities such as canSetValue, canPress, canFocus, canScroll, and adjust.
• Stable window refs from list_windows, with explicit targeting such as screenshot({ window: "@w1" }) and click({ window: "@w1", ref: "@eN" }).
• State IDs for stale-action detection.
• Deterministic window layout through arrange_window presets or explicit frames.
• Optional screenshot attachment mode with image: "auto" | "always" | "never".
• Ref-first actions such as click({ ref: "@eN" }), scroll({ ref: "@eN" }), and set_text({ ref: "@eN", text }).
• Batched actions through computer_actions, with one post-action semantic state update plus per-action execution metadata.
• Execution metadata that reports stealth for background-safe AX paths and default for focus/raw-event fallbacks.
• Full pointer and keyboard primitive coverage for common GUI flows, with AX-first equivalents where available.
• Browser-aware targeting, including isolated browser window preference where appropriate.
• Optional strict AX mode for background-safe operation without foreground focus, raw pointer events, raw keyboard events, or cursor takeover.
• Official QA benchmark harness in benchmarks/.

Examples

Prefer AX refs over coordinates when a matching target exists:

click({ ref: "@e1" })
scroll({ ref: "@e3", scrollY: 600 })

Use coordinates from the latest screenshot only when there is no suitable AX target:

click({ x: 320, y: 180, stateId: "..." })

Replace text through AX value semantics:

set_text({ ref: "@e2", text: "https://example.com" })
keypress({ keys: ["Enter"] })

Batch obvious actions when no intermediate inspection is needed:

computer_actions({
  stateId: "...",
  actions: [
    { type: "click", ref: "@e1" },
    { type: "set_text", ref: "@e2", text: "https://example.com" },
    { type: "keypress", keys: ["Enter"] }
  ]
})

See docs/usage.md for the full workflow and tool patterns.

How It Works

pi-computer-use has three pieces:

• The Pi extension in extensions/computer-use.ts registers the public tools and /computer-use command.
• The TypeScript bridge in src/bridge.ts manages the current window, capture IDs, AX refs, fallback policy, batching, and execution metadata.
• The native Swift helper in native/macos/bridge.swift talks to macOS Accessibility, AppKit, and CoreGraphics. On macOS 14+, the modern helper also uses ScreenCaptureKit for window capture.

The result is semantic-first GUI control: Pi sees useful AX targets first, falls back to screenshots only when needed, and reports whether each action stayed background-safe.

Documentation

• Usage guide: tool workflow, AX refs, text input, browser flows, batching, and strict AX mode.
• Configuration: config files, environment overrides, browser control, and stealth mode.
• Development: local setup, helper builds, validation, release signing notes, and PR workflow.
• Troubleshooting: permissions, helper setup, stale refs, browser refusal, and strict mode errors.
• Benchmarks: benchmark commands, metrics, regression policy, and local comparison workflow.
• Contributing: issue-first contribution rules and PR checklist.

Development & Benchmarks

Install dependencies:

npm install

Run checks:

npm test

Run the local checkout in Pi without loading another installed copy:

pi --no-extensions -e .

Run the default QA benchmark:

npm run benchmark:qa

Run the wider benchmark that may open apps:

npm run benchmark:qa:full

Release & Install Notes

The package is published on npm as @injaneity/pi-computer-use.

npm install @injaneity/pi-computer-use
npm install @injaneity/pi-computer-use@0.2.5

Pi installs should pin a GitHub release tag:

pi install git:github.com/injaneity/pi-computer-use@v0.2.5
pi install -l git:github.com/injaneity/pi-computer-use@v0.2.5
pi install /absolute/path/to/pi-computer-use

Remove:

pi remove git:github.com/injaneity/pi-computer-use@v0.2.5
npm remove @injaneity/pi-computer-use

For a different release, replace v0.2.5 or 0.2.5 with the version you want to pin.

Screenshots

License

MIT

See Also

• Pi
• @earendil-works/pi-coding-agent
• @earendil-works/pi-ai