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 Chat tab to read the conversation and tool activity. To return to an earlier conversation, open Chat history, the icon beside Chat, or click the conversation’s title at the top of the panel. Select New (+) to start a separate conversation.
In Chat history, point to a conversation and select Remove chat (trash icon) 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.
To reference a file, type @ followed by part of its name in the desktop
composer. For example, type @header to find header.pen under Open editors
or Workspace files. Highlight the file and press Enter to insert a
reference to its path. To remove the reference before sending, select it in
the composer and press Backspace.
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.
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.
- Type
/or$in the desktop composer, or select Add to context (+) → Pick a skill - Choose a skill from the menu
- 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
Select the factory icon in the composer to open Run settings. Under Parallel agents, choose an agent count from one to six and a mode:
- Divide and conquer assigns different parts of the task to different agents.
- Build separate versions gives each agent the same task so you can compare their results.
In Build separate versions, 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 2 and Build separate versions, 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 Chat tab to read each agent’s conversation.
Under Explore variations, choose how many variations to make after the first result. Off stops after the first design. Choose 1 to make one more design, or select up to 5. Each variation builds on the previous result.
- Layout explores a new arrangement using the same visual style.
- Style explores a different visual direction.
- Creative draws inspiration from another field and can rethink the concept or interaction as well as the look.
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.