dsh-browser

Getting an AI assistant in DeepSeek Harness to navigate a page, fill a form, or export a PDF has usually meant hand-assembling a local browser driver, a Node.js SDK, or even a Rust toolchain, especially across platforms. dsh-browser removes that friction: as a lightweight DSH Bundle it registers 31 native browser tools for the model with a single install command, requiring no local browser, Rust runtime, or third-party SDK.

On the first tool call the Bundle selects the asset matching the current OS and CPU architecture, downloads the pinned browser-cli release and its SHA256SUMS from a fixed HTTPS origin, then verifies the digest, executable format, target architecture, and CLI version before atomically installing the binary into a user cache directory. Every subsequent call and restart reuses that verified cache, so a populated cache works fully offline. The tool surface spans diagnostics and authentication, session and context management, URL navigation, wait/click/fill interactions, page snapshots, screenshots, PDF and file downloads, arbitrary page JavaScript execution, and raw CDP commands, covering everything from simple queries to complex automation. Screenshots are persisted through the DSH image-attachment service, while PDFs and downloads write directly to the requested host path.

Security is built in end to end: model input is passed as a child-process argument array with shell set to false to eliminate command-concatenation risk; returned data and diagnostics automatically strip API keys, Authorization headers, and Chrome DevTools WebSocket URLs; downloads use a fixed HTTPS origin with mandatory digest verification. The plugin currently supports Windows x64, macOS Apple Silicon, and Linux x64, and is well suited for Harness users who want their AI assistant to drive a cloud browser directly without pulling in a heavyweight runtime.