Drawgent Puts Claude Code on a Live Whiteboard
Write AGENT: make this a sequence diagram on the canvas, and the drawing changes in front of you. Drawgent, a Show HN from September 26, hooks your own Claude Code, Codex or opencode agent into a live Excalidraw board. You can talk to it in a chat panel, or you can just scribble instructions directly on the canvas next to the thing you want changed.
Under the hood it is a single Rust binary that serves an Excalidraw editor and keeps the scene state, then bridges your agent in over ACP. The agent gets an MCP server with tools like get_scene, get_screenshot and add_elements, and the screenshot tool renders through headless Chrome, so the agent can actually look at what it drew instead of guessing from JSON. It also works with excalidraw.com rooms, end-to-end encryption included, and a setup wizard checks that your CLI is installed and logged in. Builds ship as static musl, Docker and Nix.
The interesting design choice is the annotation channel. Most agent UIs make you describe a location in words. Here the instruction sits spatially next to the object it refers to, which is how people actually mark up a diagram for a colleague. That is a cheap trick with a big effect on ambiguity.
It is small and early, and there is no explicit license in the repo yet, so check before you build on it. But bring-your-own-agent into a visual workspace is the right pattern. Your subscription, your harness, somebody else's canvas. Project: tangled.org/yanndegat.tngl.sh/drawgent.
← Back to all articles
Under the hood it is a single Rust binary that serves an Excalidraw editor and keeps the scene state, then bridges your agent in over ACP. The agent gets an MCP server with tools like get_scene, get_screenshot and add_elements, and the screenshot tool renders through headless Chrome, so the agent can actually look at what it drew instead of guessing from JSON. It also works with excalidraw.com rooms, end-to-end encryption included, and a setup wizard checks that your CLI is installed and logged in. Builds ship as static musl, Docker and Nix.
The interesting design choice is the annotation channel. Most agent UIs make you describe a location in words. Here the instruction sits spatially next to the object it refers to, which is how people actually mark up a diagram for a colleague. That is a cheap trick with a big effect on ambiguity.
It is small and early, and there is no explicit license in the repo yet, so check before you build on it. But bring-your-own-agent into a visual workspace is the right pattern. Your subscription, your harness, somebody else's canvas. Project: tangled.org/yanndegat.tngl.sh/drawgent.
Comments