Drawing from an agent

Calibro Sketch ships its own Rovo agent, and the same tools are in Claude. Describe a system — draw the payments platform: browser clients behind a load balancer, a Node API, Postgres, Redis, and two workers consuming from Kafka — and the agent draws it as a real, editable Sketch in the space you name. Then keep talking: add a CDN in front of the clients, what does this diagram show?, put it on the Payments architecture page.

The agent does the drawing itself, so a diagram drawn this way does not use the editor’s own describe it, draw it AI. It writes ordinary Excalidraw elements, every label bound to its shape and every arrow bound at both ends, exactly what you would get by drawing it by hand — open it in the editor and move anything.

In Rovo

Open Ask Rovo in Confluence, use the agent switcher at the top of the panel, search for Calibro Sketch Agent and pick it. Rovo’s default agent does not carry these tools; ours does. Before anything is written, Rovo shows what it is about to send and waits for you to confirm.

Teams that build their own agents in Rovo Studio can add the same ten tools under Tools → Add tools → Connected apps and write their own instructions around them — an agent that keeps a service map current, say.

In Claude and Claude Code

Calibro Sketch is listed in Claude’s connector directory. The same ten tools, run as you, in the same spaces — so the session in which Claude read your repository can be the one that draws its architecture into Confluence, and the next one revises it in place.

1. Once per site, an admin allows it. In Atlassian Administration, open Settings → Connected apps, open Calibro Sketch and turn on Allow external MCP access. The switch is per app. If it isn’t there, approve Calibro Sketch’s pending update first: 5.0.0 asks for one new permission, writing to pages, which only the embed tool uses.
2. Add the connector. Open Calibro Sketch in Claude’s directory and choose Connect — or, in Claude, open Settings → Connectors and browse for it. On a Team or Enterprise plan, an owner may need to add it for the organization first.
3. Sign in and pick the site. Atlassian asks which site to connect. Choose the one Calibro Sketch is installed on, and accept.
4. Ask. Which Sketches are in the ENG space? is a safe first question: it only reads.

Claude Code signed in with the same Claude account picks the connector up with nothing more to do, and so does Claude Desktop.

Two differences from Rovo. There is no review card: Claude asks your permission before calling a tool, unless you have told it to always allow that one — worth keeping for Embed a Sketch and Delete a Sketch. And the Calibro Sketch Agent’s own instructions do not travel with the tools: Claude works from the tool descriptions alone, which carry the essentials — read the scene guide before drawing, read a Sketch before changing it, confirm the page before embedding and the Sketch before deleting.

The ten tools

List spaces, List Sketches

The spaces you can see, and the Sketches in one of them, newest first. How the agent finds a space key or a Sketch.

Describe a Sketch

What a diagram shows, as components and the relationships between them, in plain terms. How the agent explains or checks a diagram without reading its drawing data.

Get a Sketch

One Sketch in full, with its version. The agent reads it before changing it.

Scene guide

The drawing format and this app’s shapes and colours for each kind of component. The agent reads it before its first drawing.

Create, Update

Draw a new Sketch in a space, or save a new version of one. An update names the version it started from and is refused if someone saved in between, so nobody’s work is overwritten.

Find pages, Embed a Sketch

Find a Confluence page by title, then add the Sketch to the end of it as a new page version. The page must be one you confirmed.

Delete a Sketch

Remove a Sketch from its space. Pages that embed it show it as missing. The agent names it and waits for you first.

Whatever the agent writes arrives where you are looking: a page embedding the Sketch redraws as it saves, the space’s list gains a new diagram without a refresh, and an editor you have open takes the new version in place — or, if you have unsaved work, tells you and leaves it alone.

What an agent cannot do

Everything runs as you: an agent can only see the spaces you can see, and only write where your Confluence permissions let you write. It cannot embed on a page you did not confirm, and it cannot overwrite a version someone else saved after it read the Sketch. ChatGPT, Codex and Cursor are not listed yet — their directories need something from Atlassian’s side first; it is on the roadmap.