AI agents and Skills

Use the desktop app’s agent composer to create designs or change the current canvas. Choose a model beside the send button. See Authentication for account and provider setup.

Use GitHub Copilot chat in VS Code or Cursor’s chat with pen.dev’s MCP integration. For terminal workflows, see the CLI reference.

Sessions and models

Open the Agent tab to read the conversation and tool activity. Use the session menu at the top of the panel to return to an earlier conversation. Select New (+) to start a separate conversation.

Use the × beside a session to remove it from history.

Choose a model from a connected provider in the model picker. Open the arrow beside a model to choose its reasoning effort. If the picker warns that switching providers will lose conversation context, include the needed instructions in your next request.

Select Stop generation to interrupt a running agent. Stopping leaves its completed canvas changes in place. Inspect the result before sending a follow-up request or using Undo.

Settings (⚙️) → Chat → Notifications controls alerts when an agent finishes while the app is in the background.

Add context

Canvas selections appear in the composer as context. Select the layers you want the agent to change, then describe the change.

Use Add to context (+) → Add image or file to attach a reference image or text file, such as a PNG, JPEG, Markdown file, or source file. You can also drop files into the composer. Check that each attachment appears before sending the request.

Review permissions and usage

When an agent shows a Permission Request, review the requested action and path or command. Allow approves that request. Deny rejects it.

For supported Claude models, Switch to Auto approves the current request and lets Claude evaluate future permission requests automatically. This setting carries over to future sessions. Turn it off with Auto permission mode in Settings (⚙️) → Agents → Anthropic Claude Code.

If a tool reports auto mode unavailable for this model, turn off Auto permission mode and resend the request. Use Allow or Deny when the permission request appears.

Settings (⚙️) → Chat → Dangerously skip permissions lets agents run without permission prompts, including file and command actions. Keep it off when you want to review those requests.

Open Context usage beside the model picker to inspect token usage. The content breakdown is an estimate. Est. cost at API rates, when shown, is an estimate rather than an invoice.

When a response shows Designed for…, expand it to inspect the reported usage.

Use skills

Skills add instructions to an agent request. MCP tools let the agent read or change the document.

1. Type / in the desktop composer, or select Add to context (+) → Pick a skill
2. Choose a skill from the menu
3. Add your request and send it

To add your own skill, choose Add SKILL.md file… in the slash menu. Select a SKILL.md file or its containing folder. The folder name becomes the skill’s name in the menu, so use a unique folder name. Read the skill before adding it: its instructions are sent with requests that use it.

Keep the skill in that location. If you edit its contents, add the same file again to reload it. To remove a custom skill from the menu, point to its row and select the trash icon. Removing it leaves the folder and SKILL.md on disk.

For an external MCP client, ask the connected agent to call read_skill() to read pen.dev’s design instructions, then read the referenced files needed for the task. The desktop slash menu’s custom skills are separate from this tool.

Generate alternatives

In Parallel agents (⚡), choose an agent count from one to six and a mode:

• Split Work asks agents to work on different parts of the task
• Side by Side asks agents to work on the same task to create alternatives

In Side by Side, each agent has a model selector. Use Reset to main model to use the main agent’s model for the other agents.

To compare two signup cards, select 2x and Side by Side, then ask for a signup card with a heading and a Join button. Inspect the results on the canvas. Use the agent-name menu in the Agent tab to read each agent’s conversation.

Let it cook generates successive variants after the first design. Choose Layout to explore layouts while preserving the visual style, or Style to explore visual styles. Set Variants to between two and six, or choose Don’t iterate for a single design.

With several agents running, select Stop generation, then Stop all. To stop only the selected agent, choose Stop current while it is running. If one agent fails, inspect its conversation and the canvas before retrying. Keep the result you want and delete unwanted frames from Layers. Additional agents and iterations make additional model requests.

If an agent cannot continue

For authentication errors or a missing model, check the provider in Settings (⚙️) → Agents and follow the connection troubleshooting steps. Then return to the conversation and resend the request.

If a tool fails, open its Show details control to inspect the error. Check the current canvas before asking the agent to retry the failed action.