dsh-opencli

The built-in web tools in dsh are retrieval-only. web_search and web_fetch always return anonymous results; they cannot reach your account following, bookmarks, or unread notifications, let alone post, like, or approve a ticket on your behalf. dsh-opencli closes that gap: through the OpenCLI daemon and a browser extension, it drives the real, logged-in browser you use everyday (Chrome or Edge), so your agent reads data and takes actions from your own perspective.

Three capability lines stand out. First, 170+ site adapters (Zhihu, Bilibili, Weibo, GitHub, 12306, and more) return structured results in a single command—no page-by-page clicking. Second, generic browser primitives (browser_open, browser_state, browser_click, browser_do) cover sites without an adapter; browser_do ships a built-in analyze-init-verify loop so the model can author an adapter on the fly. Third, a distilled catalogue of 1,276 commands is injected into the system prompt, letting the agent pick the right one on the very first turn without parsing multi-megabyte list output.

Safety is built in. Every write command (posting, ordering, commenting) passes through the native dsh approval gate by default—a dialog appears and nothing executes until you click Allow. Commands with unclear privilege are treated as write (fail-closed). browser_do only permits read-only and authoring sub-commands. All output is truncated and stripped of noise such as update banners to protect context.

Who it is for: dsh users who want their agent to graduate from searching to acting, especially when login-state data or controlled write operations matter. Existing OpenCLI users also benefit—the plugin adds tool-level parameter validation, approval interception, and a visual dashboard (daemon diagnostics, login-state checks, adapter toggles). Headless servers with no real browser window are not supported.