Drawgent is a developer tool that does one specific thing well: it connects your existing coding agent (Claude Code, Codex, or opencode) to an Excalidraw whiteboard, letting the agent see and edit diagrams live. You type instructions in a chat panel or write AGENT: next to a shape on the canvas, and the agent takes a screenshot, reads the scene JSON, makes edits, and marks the note DONE. The agent is never bundled — it uses your install, your login, your config, your repo. The setup story is unusually disciplined. drawgent setup claude runs a preflight check that validates everything — CLI on PATH, auth status, ACP bridge installation, headless Chrome availability — and fails with the exact fix when something is missing. The ACP adapters for Claude Code and Codex are installed once (~60 MB) into ~/.cache/drawgent/adapters, reusing your existing agent binaries. For Chrome, it offers to download Chrome Headless Shell (~120 MB, no sudo) or install Chromium via your system package manager. The tool refuses to start until setup has fully succeeded. The interaction model has three layers. Chat panel on the right side for direct requests. AGENT: annotations on the canvas that fire ~2.5 seconds after you stop typing, with positional context about what the note points at and what's nearby. And laser zones — use Excalidraw's laser tool to circle part of a diagram, and the chat panel opens with a chip listing what's under the zone, ready for your instruction. Strokes within 4 seconds form one zone. It's a thoughtful interaction vocabulary for spatial AI collaboration. Room support is the multiplayer angle. drawgent up --room joins an excalidraw.com collaboration room as a bot participant with its own cursor. Human collaborators stay on excalidraw.com, write AGENT: notes, and see the agent's edits appear live. Traffic is end-to-end encrypted with the room key. The local editor mirrors the room and retains the chat panel. The architecture is 78% Rust with a JavaScript/React frontend for the editor and renderer pages. The MCP tool surface is well-scoped: getscene, getscreenshot (with zoom via elementids), addelements, addmermaid (auto-layout), updateelements, deleteelements, and instruction management. There's a REST API for programmatic access and WebSocket sync for the browser. Docker support runs a canvas server without agents for shared hosting. Limits are stated plainly: the renderer requires Chrome (native renderer planned), Claude attach is a fork rather than injection into a running session, Codex live attach is implemented but untested against a logged-in instance, and it's one scene per workspace with no image/file sync. The codebase has 9 stars, 1 fork, and sits on Tangled (an AT Protocol-based Git hosting platform), which is itself a notable choice. This is infrastructure-grade glue for a workflow that didn't previously exist as a coherent product. The bet is that spatial, visual collaboration with coding agents is a real interaction paradigm, not just a demo. Whether that bet pays off depends on whether developers actually think in diagrams when working with agents — and whether the ACP protocol stabilizes enough to make these bridges durable.
