Initial public release.
What's inside
- Selection context sync — every browser tab publishes a bounded snapshot of the selected node over a same-origin WebSocket into an in-memory page registry; the active page follows window focus and carries an on-screen
AI-XXXXbadge. - HTTP API —
GET /ai-assistant/context(context envelope),POST /ai-assistant/context(anonymous snapshot channel for scripts and debugging). - In-process MCP endpoint —
POST /mcp, stateless Streamable HTTP subset; tools:get_selection,set_widget_text. Verified live with Claude Code and OpenCode. - One narrow write surface —
set_widget_textinserts text into a text-like widget of the selected node on the active page, guarded by revision and page-identity checks plus size bounds; on the canvas it behaves like a manual edit (undo works). - Zero runtime dependencies; read-only apart from the guarded write; RAM-only storage, no telemetry, no outbound requests, no payload logging.
Install
cd <ComfyUI>/custom_nodes
git clone https://github.com/alksef/ComfyUI-AIAssistant.gitRestart ComfyUI — see the README for MCP client configuration (Claude Code, OpenCode, curl).
Verified against ComfyUI 0.33.3, frontend 1.50.6, Python 3.10–3.13.