← Guide

Beta test: connect your own agent

Thank you for testing. There is one thing we need answered right now, and it is not whether the website works: it is whether the agent you already use can build a real game here, starting from one sentence. Everything below is aimed at that.

The test plan

Ten tests, in the order they build on each other. Each one says what to do and what we are trying to learn — so you can tell us not just whether it worked, but whether it worked well.

1. T1 Connect, and disconnect

   1. Issue a token on the connect page, name it after the machine you are on.
   2. Run the one command it gives you in your agent’s terminal, then start a new session.
   3. Ask your agent: “what can you do on Lumen Play?”
   4. Come back and revoke the token. Ask it again.

   That connecting is one command and one paste, that your agent can say what it now has access to without you explaining it, and that revoking really does stop it. Tell us how long the whole thing took and where you hesitated.

2. T2 A cold start — give it nothing

   1. Open a brand-new session, with no instructions from you about how this platform works.
   2. Say one sentence: the game you want. “Make me a small game about tides.”
   3. Watch what it reaches for first, and let it finish without helping.

   Whether your agent ARRIVES knowing the platform or discovers it one mistake at a time. This is the single most important thing in the beta. If it invented something that does not exist, or built the game out of the wrong pieces, that is the report we want most — paste what it tried.

3. T3 Change your mind in the middle

   1. Half-way through, ask for something different: a new rule, a new ending, a piece moved.
   2. Then ask for something that contradicts what it already built.

   That an edit lands on the game that exists rather than starting a new one, and that a contradiction produces a question or a clean refusal instead of a broken world.

4. T4 Find the ceiling

   1. Keep going on one game. Bigger, then more complicated: more objects, more rules, more states, a longer chain of things that trigger each other.
   2. Push until something gets worse — slower, confused, wrong, or refused.
   3. Write down where that was and what it looked like.

   We are not going to tell you where the ceiling is — we want to know where YOU hit one, and whether it announced itself or just quietly degraded. A ceiling that says so is a feature; a ceiling that lets you keep going while getting worse is a bug. Both are worth reporting; the second is urgent.

5. T5 Break it on purpose, then go back

   1. Ask for something you know is a bad idea, and let it apply.
   2. Ask your agent to list the versions of the game and put it back to the one before.
   3. Play it and confirm the bad turn is gone.

   That a bad turn is never a lost game. Tell us if the version list was hard to read, or if restoring took more than one instruction.

6. T6 Make it look at its own work

   1. Ask your agent to look at the game it just built, and to fix what it sees.
   2. Then look yourself, and compare.

   Whether what it says it sees matches what you see. Disagreements here are gold — send us the picture and its description of it.

7. T7 Take it live

   1. Ask your agent to publish the game.
   2. Find it yourself in the feed, open its page, and play it through to the end.
   3. Play it once more on your phone.

   That publishing from your own agent produces a game a stranger can find and finish — and that it holds up on a phone. Anything that looks wrong on the page or the card is worth a screenshot.

8. T8 Give the library something

   1. Ask your agent to design a new object, modular or frame and publish it to the shared library.
   2. Open the library and find it.
   3. Ask it to build a second game that uses the thing it just made.

   That what one creator makes is usable by the next one, and that your agent can find its own contribution again afterwards.

9. T9 Try to make it misbehave

   1. Issue a second token as read-only, connect with that, and ask it to change something.
   2. Ask it for another creator’s draft.
   3. Ask it to publish something you have not looked at.

   That “read-only” means it, that your account is the only account it can touch, and that nothing goes public without a deliberate act. If any of these three succeed, stop and tell us immediately — that one is not a normal report.

10. T10 Bring your own agent

    1. If you use something other than Claude Code — Cursor, a client of your own, anything that speaks MCP — connect that instead and repeat test 2.

    Which clients work without special handling. This is the one test we cannot run ourselves for every tool, and the answer decides what we support.