Connect your AI
Yumina has an MCP server at https://yumina.io/mcp. Add it to the AI tool you already use, sign in once, and that AI can open your cards, write lore, variables, behaviors and interface code, make pictures and playtest, with every change appearing live on your canvas. It cannot publish. That click stays yours.
The Creation assistant inside Studio does the same work with no setup. Connect your own AI when you want your own model, your own subscription, or an agent that keeps working from a terminal.
Quick start
- Open the ⋮ menu at the top right of the canvas and click Connect your AI. The card has to be saved once first.
- Pick your tool. The dialog shows its steps, and they all come down to the address above.
- Click Allow on the Yumina page that opens.
- Send your AI a first message. The dialog notices the connection and says so.
Use Yumina to open my card "Mushroom Tavern" and tell me what's in it right now.Per client
Claude
Desktop or web. Open Settings → Connectors → Add custom connector, name it Yumina, paste the address, click Continue, then Add. Click Connect, then Allow on the Yumina page.
ChatGPT and Codex
Settings → Plugins → Add → Add MCP server. Name it yumina, pick Streamable HTTP, paste the address, Save. Sign in to Yumina and click Allow.
Cursor
Click Add to Cursor in the dialog, or add this to .cursor/mcp.json, then sign in to Yumina and click Allow:
{ "mcpServers": { "yumina": { "url": "https://yumina.io/mcp" } } }Claude Code
claude mcp add --transport http yumina https://yumina.io/mcpThen type /mcp in Claude Code, pick yumina, Authenticate, and click Allow in the browser.
Codex CLI
codex mcp add yumina --url https://yumina.io/mcp
codex mcp login yuminaThen open ~/.codex/config.toml and add one line under [mcp_servers.yumina], because Codex gives a tool 60 seconds by default and a playtest or a picture takes longer:
tool_timeout_sec = 300DeepSeek Harness
It cannot sign in yet, so its tab in the dialog makes a token for the card you have open, gives you a yumina.yml to save, and a command to run:
npx @deepseek-ai/dsh web --patch yumina.ymlThe token reaches that one card only and expires after 90 days.
Any other tool
If it cannot sign in, open Can't sign in? Make a token for this card in the dialog. Put the address and the Authorization header it shows into your tool. The token is shown once.
Example usage
Add a sly merchant called Vex who talks in riddles and has trust issues.Make a hunger variable that drops over time. When it hits zero, the player dies.Build a dark fantasy interface with health, mana and inventory slots.Go through the sticky notes on my card and make every change they ask for.Playtest five turns as a cautious player and tell me which variables never moved.Check the card for anything wired wrong and fix it.Available tools
| Job | Tools |
|---|---|
| Find your cards | list_my_cards, create_card |
| Read a card | get_world, read_entities, grep_world, list_assets, read_ui_doc, load_skill |
| Check it, for free | validate_world, analyze_token_cost, simulate_behaviors |
| Write the story | write_entry, write_variable, write_behavior, write_worldbook, write_audio, write_scene_image, write_note, update_settings, set_card_name, delete_entities |
| Build the interface | write_custom_ui, edit_custom_ui, edit_ui_doc, set_ui_knobs |
| Try it | playtest, up to 10 moves at a time, and screenshot_ui |
| Make pictures | generate_image, set_cover |
Most tools take a variant, so it can work on a card's other language versions. It reads your sticky notes and knows where each one is stuck. The server tells every AI to start with get_world, and delete_entities is marked destructive, so clients that respect that hint ask you first.
What it costs
Reading, writing and checking are free. Playtest moves, screenshots with moves, and pictures cost mushies at the normal price, the same as when you click them yourself, and failures are refunded.
Security
- No password in any config file. Sign-in is OAuth 2.1 with PKCE. The AI holds a short-lived token with the "cards" permission and nothing else, and the consent page lists exactly what that covers.
- It cannot publish or touch your account. On a published card, its edits wait as unpublished changes until you submit them. No balance, password or settings.
- A checkpoint before every change. ⋮ → Change History shows "Before: Claude", or whichever AI it was, and you can roll back to it. At most one per AI every 15 minutes.
- Limits catch a runaway agent. 30 playtest turns, 20 pictures and 30 screenshots an hour, and 20 new cards a day.
You should still:
- Read what it wrote before you publish, the same as with the Creation assistant.
- Disconnect what you no longer use. Connected AIs in the dialog, or Settings → Account → Connected apps, lists each one with when it was last used. Disconnect and it loses access that second. Card tokens are listed and revoked the same way.
- Switch accounts cleanly. Disconnect, sign out of Yumina in your browser, sign in to the other account, and connect again. For Codex CLI run
codex mcp logout yumina, thencodex mcp login yumina.
Without a connector
If your tool has no MCP support at all, give it the world file format instead. It writes the card's JSON and you import it.
