browser-auto

A real-browser automation plugin for DeepSeek Harness (DSH), built as a dynamic Cordis plugin. It turns a genuine headless Edge/Chrome instance into a first-class capability of the agent: 12 browser_* model tools for opening pages, reading structured snapshots, clicking, typing, pressing keys, navigating history, taking screenshots, waiting for rendering, and arbitrary in-page evaluation — plus a live screenshot panel rendered inside the GUI so you can watch exactly what the browser sees.

No npm dependencies. The driver talks to the browser over the raw Chrome DevTools Protocol using Node's built-in fetch and WebSocket (Node ≥ 22).

Highlights

• Real browser, real input — every action goes through CDP: real mouse events at element centers, trusted Input.insertText keystrokes that work with React/Vue-managed inputs (which ignore synthetic value setters), and a key map for Enter/Tab/arrows/PageUp/…
• Structured snapshots, not raw HTML — each page digest returns the URL, title, body text, an indexed input list and an indexed clickable-element list, so the agent locates targets by snapshot index, CSS selector, or visible text.
• Zero dependencies — only Node ≥ 22 built-ins (fetch, WebSocket, child_process); no npm install, no Playwright or Puppeteer.
• Live GUI panel — the plugin's run card shows status, URL/title, a screenshot refreshed every 2 seconds, and back / reload / screenshot / close buttons.
• Privacy-safe by default — every launch uses a fresh ephemeral browser profile that is deleted on close; no cookies or history persist between sessions; screenshots stay on your machine and are gitignored.
• Self-healing — if the browser or the driver process dies, the next tool call restarts it automatically.

What it does

DSH agent ──browser_open/click/type…──▶ host half ──stdio JSON──▶ driver.mjs ──CDP──▶ Edge/Chrome
                                                │                        │
                                                └─ /dsh-browser/shot.png ┘
                                                │
DSH GUI ◀──browser-state/action RPC── client half (panel)

Requirements

• DeepSeek Harness with the dynamic-Cordis extension (the cordis_define / cordis_run tools)
• Node.js ≥ 22 (global fetch + WebSocket)
• Microsoft Edge or Google Chrome. Windows default install paths are probed; edit BROWSER_CANDIDATES in driver.mjs for other platforms.

Install

1. Clone / place this repo somewhere on the machine, e.g. C:\browser-auto.

2. In host-half.js, set the DRIVER constant to the absolute path of driver.mjs on your machine:

   const DRIVER = 'C:\\browser-auto\\driver.mjs'

3. In your DSH session, define the plugin — paste the entire content of host-half.js into code.host and of client-half.js into code.client:

   cordis_define(plugin: { kind: "new", idPrefix: "brws" },
                name: "browser-auto",
                purpose: "Real browser automation: browser_* tools + live screenshot panel",
                code: { host: <host-half.js>, client: <client-half.js> })

4. Run it and authorize the Client half in the GUI (single check mark on the run card):

   cordis_run(pluginId, packageId, mode: "run")

5. Smoke-test the driver standalone any time:

   node driver.mjs --selftest

Tools

Privacy & safety

• Ephemeral profile: every browser launch uses a fresh throwaway Edge/Chrome profile directory (profile-<id> beside the driver), removed when the browser closes. No cookies or history persist between sessions.
• Everything runs locally: pages, screenshots, and profiles stay on your machine; there is no telemetry or remote control.
• Screenshots are written only to <driver dir>/shots/ and are served to the GUI over the local DSH origin. shots/, profile-*, and the runtime directory are gitignored.
• The model tools can navigate arbitrary sites and run browser_eval — treat this like giving the agent a real browser. Sites with bot detection (e.g. Baidu's slider CAPTCHA) may block headless browsers; that is the site's policy, not a plugin bug.

Known limitations

• One tab, top-frame DOM only (no cross-frame automation).
• Default BROWSER_CANDIDATES are Windows paths; on macOS/Linux add your browser path and adjust the --headless=new flag if needed.
• The dynamic host sandbox has no process/env, so the driver path must be a literal in host-half.js.
• Node ≥ 24 (V8 13) rejects single-line if (x) a() else b() — keep braces or newlines in plugin code.

License

MIT