Computer Use for DeepSeek Harness on macOS

简体中文 · Capabilities · Security

An independent community plugin with a familiar app/window API, native macOS accessibility and screenshots, and opt-in Chromium DOM automation. Native execution uses the official MIT-licensed DeepSeek Harness CUA provider; this project adds an original compatibility facade and a browser adapter. It contains no OpenAI implementation code and is not affiliated with OpenAI or DeepSeek.

This is not a complete clone of Codex's private runtime. Locked-desktop control, OpenAI's browser extension and signed-in profile integration, Appshots, browser-history authorization, and WebMCP are not supplied. See the capability matrix before installation.

Requirements

• macOS 14 or newer, Apple Silicon or Intel; an unlocked desktop.
• Node.js 24.9+ and DeepSeek Harness 0.2.2-rc.1 (the tested prerelease).
• A model route declaring image input for screenshots.
• Accessibility and Screen Recording permissions for the official DSH Computer Use.app, followed by Harness's separate app-consent prompts.
• Optional browser control: a browser you explicitly started with a loopback CDP listener and a dedicated profile. Browser control is disabled until its endpoint is configured.

The official provider dependency ships signed, precompiled native helpers. This plugin neither recompiles nor re-signs them. Older Harness 0.2.1 profiles are intentionally excluded because their CUA contracts differ.

Install

From a reviewed checkout or a release tarball:

dsh plugin --profile desktop add ./dsh-computer-use-macos-0.1.0.tgz

Or install a reviewed Git commit (replace <commit> with the release's full SHA):

dsh plugin --profile desktop add github:yw5579358/dsh-computer-use-macos#<commit>

Built lib/ files are committed, so Git installs do not require a prepare/install script. Enable only one computer-use provider per profile; remove a competing provider before adding this bundle. Restart Harness when the host does not support hot reload.

Use a PTC session. The bundle mounts the native provider with agentPreset: ptc and retains the official authorization flow. This plugin does not silently switch existing sessions or their model. A GUI host that supports the CUA Client page exposes its permission controls; source/headless users must complete TCC permissions on the Mac itself.

Discover index

For Harness's experimental Custom plugin indexes, add this explicitly trusted source:

https://raw.githubusercontent.com/yw5579358/dsh-computer-use-macos/codex/initial-release/plugin-index.json

The format-1 index pins the tested plugin source to an immutable commit. This is a self-hosted community index, not an official curated listing.

Use native apps

In a persistent PTC execution:

const entry = await tools.computer_use_bootstrap({})
const { connectComputerUse } = await import(entry.nativeBridgeUrl)
const cua = connectComputerUse(() => tools.cua, {
  text: value => console.log(value),
})
const app = await cua.getApp('Calculator')
// Initial AX state is emitted automatically. Use an element from that observation.
await app.click(elementIndex)
await app.getAXState()

The getter resolves the current execution's tools.cua every time. Never capture an expired execution's binding. The Host continues to own approval, cancellation, logging, target identity, image admission and saved application consent.

Methods include app discovery, explicit window selection, AX observations, screenshots, click/hover/drag, value replacement, literal text selection, advertised secondary actions, Unicode input, shortcuts, and plain/HTML clipboard paste with the upstream clipboard-preservation semantics.

Important differences from the Codex interface:

• Screenshot results retain {bytes, mimeType, pixelWidth, pixelHeight, windowId}; the image is already admitted by the Host. Do not print Base64 or upload it a second time.
• scroll({x, y, deltaLinesX, deltaLinesY}) uses bounded line deltas. It does not accept page counts or an AX scroll-target index.
• click(index, {key:'super', durationMs:800}) maps to native modifiers and a bounded click hold.
• AX reads are full observations, not Codex's incremental diff format.
• Keyboard/paste require the intended focus. Raw Unicode HTML has an AppKit encoding limitation in the tested text field; prefer plain text. See the validation report.
• After an input action, including a failed one, observe again before further input. Do not automatically retry an uncertain action.

Optional browser tool

Browser configuration and examples are in the browser guide. The supported model path is computer_browser through the Harness tool pipeline, with declarative locator operations. Access is opt-in, profile-wide and bounded by Harness's normal tool policy; it is not per-domain authorization. CDP can access all tabs in the configured browser, so use an isolated profile.

The lower-level ./browser export is a developer library for testing or embedding, not a substitute for the Host tool approval pipeline.

Test and develop

npm ci --ignore-scripts
npm run check
npm run test:browser
npm pack

The browser integration test starts an owned headless Chrome/Chromium process with a temporary profile and a local fixture page. Set DSH_TEST_CHROME when the standard macOS Chrome executable is unavailable. It never uses your daily browser profile. Native desktop testing requires explicit system grants and an owned fixture; a mock test does not establish real desktop behavior.

See validation results, research sources, contribution instructions, and third-party notices.

Remove

dsh plugin --profile desktop remove dsh-computer-use-macos

Removal does not revoke macOS TCC permissions or erase upstream Harness application grants. Revoke those separately in System Settings and the Harness CUA permission page. No tokens, personal screenshots, session logs or browser profiles belong in this repository.