Browser Live

Description

A visible Agent browser (in-house, zero external dependencies): 21 browser_* tools directly drive local Chrome/Edge (CDP, including human-like mouse trajectories). The live observation window supports both an embedded panel and a standalone webpage (/bl/view, suitable for a secondary display/fullscreen), with mouse and keyboard takeover, proxy configuration, and additional launch arguments. Three-part webpage understanding toolkit: browser_read (four-level content fallback + Markdown + paragraph-aware continuation), browser_scrape (structured item+fields extraction), and browser_search (search in a real browser; engine-agnostic: preset bing/baidu/ddg plus any search page via engineSpec; empty results are triaged by four causes). The complete invocation scheme is packaged as the skill skills/browser-automation (registered by index.js; POST /bl/skills/reload rescans without restarting). Optional takeover of your everyday browser: MV3 extension + local WS bridge, per-site authorization, read-only by default, and clicking/typing enabled only after opening “Allow actions.” Chrome and Edge can be connected simultaneously; specify which one with use for each call. If the preferred launch fails, it automatically falls back and retries with --in-process-gpu. It may close only tabs opened by the agent itself; tabs you opened manually are never touched. Every agent step is persisted as an audit trail (one append-only JSONL file per day under audit/ + hash chain, with sensitive inputs redacted; verify-audit-chain validates the chain). Login state is persisted in $DSH_HOME/dsh-browser-live. When moving to another machine, note: taking over “your everyday browser” requires manually installing the extension, pasting the token, and authorizing sites individually; the agent can only prepare the environment (browser_ext_setup). The “plugin-provided instance” option requires no authorization and is fully automatic. See README “Moving to another machine: portability and steps that must be manual” for details. Inspired by dsh-ego-browser (MIT); see NOTICE for the full distilled sources and license text.