Open a tool as a canvas, then use ⌨ Terminal - build. The app starts a fresh Claude Code session with the current tool's context and an MCP connection to the live window.
1You open a toolThe Tauri window navigates directly to the tool's local URL.
2The bridge observesConsole, errors, network, DOM, navigation and performance events enter the stream.
3The agent looks and actsMCP exposes the window's pixels, DOM, JavaScript and deliberate UI actions.
4The agent iteratesEdit local source, update the tool, reload and inspect the result.
The terminal launcher assembles the Hive knowledge prompt, current canvas identity, per-canvas MCP config, allowed tool list and directory grants. A local tool's editable source directory is granted; an installed team tool is treated as a pinned copy that must be changed through its team repository.
screenshotinspectevalactionlist_windows
The surface stream is read-only. Clicking, filling or selecting happens only through an explicit agent or human action. The canvas prompt tells the agent to observe first and never drive the UI unprompted.
Hive is not a chat window.
The built-in chat was removed. The adapters, CLI and local APIs can be used by Kiro, Q, Claude Code or another harness that can invoke them. The one-click canvas build launcher currently starts Claude Code; equivalent context injection for other harnesses is not automatic today.