# Global Agent Chat example (work in progress) This browser sandbox plugin provides a custom Chat view using the existing host-owned Agent runtime. Choose a configured Catty or Claude Code backend, explicitly choose connected terminal targets, open a new or owned conversation, then Send. Stop cancels the exact turn. Refresh and reselection are explicit; expired authority never silently re-shares targets or retries inference. The command runtime remembers this sample's current opaque chat handle across modal closes. A later explicit Open releases that exact lease before opening the selected conversation. View pagehide cleanup is best effort; core source, expiry and duplicate-open checks still apply. No handle is stored persistently. Core stores source-isolated conversation history. The plugin receives text, usage and safe activity summaries, never native session IDs, credentials, raw tool arguments/results or executable/MCP configuration. Tool actions use the plugin's `ai.agent.actions` Allow / Ask / Deny setting. Other backends report unsupported/unavailable until their native tool boundaries are verified. Build with `node examples/plugins/global-agent/build.mjs`, then package the `.package` directory using `@netcatty/plugin-cli`. Install the resulting `.ncpkg` through Settings → Plugins. This is a minimal example, not a background service. Real provider/Claude usage requires a user's already configured backend and separate approval for paid acceptance testing. Automated tests use synthetic fixtures and do not establish authenticated Claude service/native acceptance. ## Current UI build The page now uses the host-rendered contract. Run `node build.mjs` in this example directory, then package `.package/` with the plugin CLI. `host.js` registers ordinary DOM views; `view.html` and `view.css` are bundled source templates. Theme values are inherited from Netcatty. There is no separate page preload or `window.netcattyView` API. See `docs/plugin-platform/host-rendered-ui.md`.