Sketch It, Ship It: Drawgent Puts Claude Code on a Live Excalidraw Whiteboard

The most interesting developer tools rarely arrive with a press release. On September 27, 2026, a small open-source project called Drawgent reached the Hacker News front page with 153 points and 41 comments — and it captures, in one binary, a question the whole industry is fumbling with: what does a shared workspace between a human and a coding agent actually look like?

Drawgent, written in Rust by developer Yann Degat, connects your own coding agent — Claude Code, Codex, or opencode — to a live Excalidraw whiteboard. Not a static “generate me a diagram” one-shot, but a persistent canvas where you sketch an architecture, scribble a note next to a shape, and watch the agent redraw it properly, check its own work, and mark the note DONE.

How it actually works

The mechanics are straightforward and refreshingly honest about their limitations. Drawgent runs a local server (default port 7300) that serves an embedded Excalidraw editor and drives your agent over ACP (Agent Client Protocol). The agent gets a set of MCP (Model Context Protocol) canvas tools: get_scene, get_screenshot (with vision and zoom on specific elements), add_elements, add_mermaid for auto-layout, update_elements, delete_elements, and a few instruction-resolution helpers. Arrows bind by element ID and route edge-to-edge; labels follow shapes when they move.

The interaction loop is the clever part. You can talk to the agent through a chat panel on the right side of the canvas — replies and tool calls stream in, and you approve permission prompts inline. Or you can skip chat entirely: write a text element starting with AGENT: next to the part of the drawing you want changed, and about 2.5 seconds after you stop typing, the agent receives your note along with its position, what it points at, and what’s nearby. It resolves the request into a green DONE: note. Edit the note back to AGENT: and you can send it again.

There’s even a laser-pointer gesture: circle a region of the diagram with Excalidraw’s laser tool, and the trace becomes a red-outlined zone. The chat panel opens with a chip listing which elements fall under it (“Laser zone · API, Redis”), and your next instruction goes to the agent together with the zone’s bounds. When the agent finishes, the trace disappears.

Your agent, your login, your repo

A deliberate design choice stands out: Drawgent never bundles an agent. It uses your install, your login, your config, and your repository. The Docker mode runs a canvas server with Chromium for rendering but explicitly excludes agents. In a year where every vendor wants you to route through their hosted everything, a tool that stays a thin bridge over software you already pay for is quietly political.

The --attach mode is where it gets practical for people already deep in agent workflows. For Claude Code, attaching forks a running session — the fork carries the full conversation and is driven by Drawgent over ACP, while your terminal session is left untouched (Claude Code has no public way to inject into a live terminal). For opencode, attach is truly live: messages flow into your running TUI session, and Drawgent injects its MCP tools into the server at runtime. Codex queues commands against existing threads and mirrors replies from the rollout file — implemented, though the README admits it hasn’t been tested against a logged-in Codex yet.

A newer feature, --diagram docs/architecture.excalidraw, keeps the drawing in your repository as a standard .excalidraw file — pretty-printed, sorted keys, no per-edit sync noise — so it diffs cleanly and commits like any other source file. If the file changes underneath (a git pull, a checkout, a hand edit), the file wins and the canvas updates live, deletions included. The commits themselves are co-authored by Claude Opus 5.5, which tells you how the tool was built.

For teams, --room joins an excalidraw.com room as a collaborator named “Agent” with a cursor that follows its edits. Humans stay on excalidraw.com; their AGENT: notes reach your local agent, and its edits appear live for everyone. Traffic is end-to-end encrypted with the room key.

Why a whiteboard matters

The HN comment thread surfaced the real debate. One camp argues the value of a diagram comes from thinking about it — the act of drawing exposes where your team’s knowledge is thin. Handing that to an agent discards the point. Another commenter noted that math papers written by models read poorly because the model doesn’t know which parts to emphasize and which to gloss over; diagrams may suffer the same flatness.

The author’s answer in the thread is telling: “as I’m very bad at making good drawings by hand, here I can sketch them in a quick and ugly manner and get them instantly properly arranged and iterate.” That’s the defensible niche — Drawgent as an equalizer for people who think in structure but draw badly, not a replacement for the thinking that diagrams are supposed to force.

Meanwhile, the thread became an impromptu survey of the agent-whiteboard space: Excalidraw’s own first-party MCP endpoint (untouched for six months, commenters noted), Mermaid-based approaches like an Obsidian relay plugin, TLDraw, whiteboard-mcp.com, and an open-sourced competitor called whiteboard-agents that appeared within hours. One commenter reported that Opus 5.5 via Claude Code can generate Excalidraw scenes with no additional software at all, purely from conversation. The space is crowding fast, which usually means the underlying need is real.

The gaps

The README is unusually candid about limits: the renderer needs Chrome (a native renderer is planned), Claude attach is a fork rather than a true injection, images and files aren’t synced, and there’s one scene per workspace. The macOS build was cross-compiled from Linux with zig and — at time of writing — “not yet tested on a real Mac.”

But as a statement of where agent UX is heading, Drawgent is more interesting than its star count suggests. The chat-panel-plus-shared-artifact pattern — a persistent, inspectable, jointly editable surface between human and agent — is a much better fit for how developers actually collaborate than the terminal-scrollback status quo. The terminal is where agents run; the whiteboard is where humans think. A bridge between the two was overdue.

Drawgent is available at tangled.org, with static Linux builds, a macOS universal binary, and Docker images for shared canvas hosting.