dsh-browser
Browser automation for DeepSeek Harness: a Playwright-powered browser_* tool family so the agent can drive real web pages — fill forms, click through flows, scrape SPA content, and take screenshots it can then view with read_image.
Install
dsh plugin --profile <name> add dsh-browserThe plugin depends on playwright-core (bundled with the package) and needs a Chromium-based browser. It tries chromium, then chrome, then msedge channels automatically; if none is found:
npx playwright install chromiumor configure launch.executablePath / launch.channel in the plugin config.
Tools
| Tool | Purpose |
|---|---|
browser_open |
Open (or reuse) the browser, optionally navigate to a URL |
browser_navigate |
Navigate the current page and wait for load |
browser_click |
Click an element by CSS selector |
browser_type |
Type text into an input (optionally press Enter) |
browser_select |
Select option(s) in a <select> |
browser_screenshot |
Capture a PNG to <workspace>/browser-screenshots/; view it with read_image |
browser_eval |
Evaluate JavaScript in the page, return JSON |
browser_get_text / browser_get_html |
Read visible text or outer HTML |
browser_wait |
Wait for slow pages / lazy content |
browser_close |
Close the browser (reopenable by any tool) |
browser_install |
Explain or verify browser availability |
The browser persists across tool calls — open once, drive many times, close when done.
Configuration
| Field | Default | Meaning |
|---|---|---|
launch.executablePath |
— | Absolute path to a browser binary (takes precedence) |
launch.channel |
auto (chromium → chrome → msedge) |
Browser channel |
launch.navigationTimeoutMs |
— | Default navigation/action timeout |
launch.viewport |
1280×800 |
Page viewport |
screenshotDir |
browser-screenshots |
Screenshot directory under the workspace |
# profile cordis.patch.yml
- id: browser
config:
launch:
channel: chrome
navigationTimeoutMs: 30000Notes
- Runs headless by default; a headed mode may come later.
- Screenshots land in the calling agent's workspace so the model can read them with
read_image.
License
MIT