Your agent’s turn.
Invite your agent to answer as you. It plays the dilemmas, creates your character sheet, and brings back a link. You get to decide how well it knows you.
Decision styles · playful beta. Unknown is not neutral; guesses do not determine traits. Labels can vary across runs and models. Play the previous edition.
Bring your agent.
Connect Align in your web chat, then let your agent read the dilemmas and create your sheet. No CLI, wallet, API key, or Align login.
ChatGPT web — setup
- Open Settings → Security and login → Developer mode. Availability depends on your account and workspace.
- Open ChatGPT Plugins, choose +, and add a connection named Align with the MCP URL above. Choose no authentication if asked.
- Check that
start_align_testandcomplete_align_testappear. Start a new conversation and enable Align in its tools menu.
Official ChatGPT instructions · Test Chat and Work separately if you use both.
Claude web — setup
- Open Customize → Connectors → + → Add custom connector.
- Name it Align, paste the MCP URL above, and leave optional OAuth fields empty.
- In a new chat, open + → Connectors and enable Align. Team and Enterprise workspaces may need an owner to add it first.
Pasting the URL into a message does not install the connector. If connection is unavailable, use the copy-and-paste fallback.
Let it speak for you.
In a conversation with Align enabled, send this invitation. Your agent answers all 12 dilemmas and brings back your sheet.
Rationales are sent to Align and shown on a public-by-link card. Unlisted does not mean private; avoid sensitive details.
Does that sound like you?
Open the returned card, read its reasons, and compare them with your own. The interesting part is where you disagree.
Test the connection with a fictional person first
Use the same prompt in either web app after enabling Align. Review its normal tool confirmation.
What counts as a successful test?
- The agent reads all 12 dilemmas and submits every answer through the tools.
- A real card opens with the complete answer record. No JSON relay, portrait charge, or homepage listing.
- Check link clickability separately: a successful API call with a broken link is only a partial pass.
Try three fresh chats per model. Record the date, app, model, Chat/Work mode, error, card URL, and manual steps. Different alignment results are not connection failures.
Other ways to play & current limitations
WebMCP · compatible agent browsers
Keep this page open in a compatible agent browser and ask it to play Align. Site tools are browser-specific, not a promise that ordinary ChatGPT or Claude web chats can use them.
checking for page-scoped site tools…
GPT Actions · ChatGPT-only experiment
A no-auth Action completed a live test with the original deck on September 4, 2026, but its editor preview did not make the result link clickable. There is no published Align GPT. Create your own private test GPT below.
- Open the GPT editor → Configure. Add a name and paste the instructions below.
- Disable Web Search and Image Generation to isolate the test.
- Create a new Action, paste the schema below, and set Authentication to None.
- Use the Action test prompt in Preview. Approve one free, unlisted card. Check the debug response and open its real card URL. Save privately if desired.
GPT instructions, schema & test prompt
Results links · experimental
Encoded answer links worked in a Claude smoke test but were not clickable in the tested ChatGPT sessions. Do not rely on them as the universal fallback. Use the JSON form below when needed.
Copy-and-paste fallback · no connector needed
Fallback: start the test with the complete prompt included:
show the full prompt
Then paste the JSON it returns below.
Copy the complete prompt, open ChatGPT or Claude, and paste it. The beta deck is too long for a reliable launch URL. Connected tools above avoid this relay.
Switch deck: reflection-12-v2.
Publishing is off by default. Your rationales are stored on the resulting public-by-link card, but they are not listed on the homepage unless you check the box.