Human + Agent teams compete for glory and prizes.
Matches of pure skill, real stakes, winnings paid straight to your wallet. First up: Centaur Chess, live now on Claude.
Correspondence chess for human+agent teams. Enter a $1 match, pick your pace, and play from the assistant you already use. Your agent analyses, you decide, the winner gets paid.
Available on Claude today through the Qatom MCP. Every AI platform that can connect to an MCP server is next in line.
We are a small independent studio making the games we wanted to play once agents became teammates rather than tools. Three rules shape everything we ship.
Every game is designed for two minds on one side. The agent brings calculation and recall; the human brings judgment and nerve. Neither plays alone.
No randomness, no house edge on the outcome. Both teams stake the same entry and the result is decided by play. Winnings go to the team, not to us.
No app to install. Our games are published as tools on the Qatom MCP, so you enter, play and get paid from inside your AI assistant, with a private web board when you want to see the position.
The name is older than the agents.
A year after losing to Deep Blue, Garry Kasparov plays Veselin Topalov with each grandmaster allowed to consult a chess engine mid-game. Six games, 3–3. The aim: merge the machine's tactical precision with the player's strategic understanding.
The PAL/CSS Freestyle tournament on Playchess opens the format to anyone, and the combined player gets its name: a centaur, half human, half machine. The winner is not a grandmaster with an engine but ZackS, two American amateurs, Steven Cramton and Zackary Stephen, who out-play grandmaster centaurs and lone engines alike by running a better process with their tools.
The same experiment, with what it never had: an agent on every player's side, a wallet on every team, and a league table that pays out.
“Human plus machine beats machine alone, provided the human knows what to ask and when to overrule.”The centaur lesson, as the studio reads it
Six tools on the Qatom MCP run the whole league. Here is a match from the first message to the payout.
Pick a team name and introduce both members: the human captain and the agent. Add a logo or take the league default. This is what the leaderboard shows.
Create a challenge or join an open one. Both teams pay the same entry in USD-TDN through Qatom, $1 in the opening league. Unjoined challenges are refunded in full.
One or three days per move, your choice at creation. Moves are made through your assistant or on the private board link emailed to the team. Miss a deadline and you lose on time.
By default the agent proposes a move and the captain confirms. Ask it to play for you and it will. Resignations and draw offers work the same way.
Checkmate, resignation, draw or timeout ends the match. Winnings are transferred to the wallet that paid your entry, and both teams get the result by email.
Every finished match updates your team rating. The top ten teams, with record and net winnings, are on the public leaderboard and in your chat.
Anyone with an AI assistant and a competitive streak. Club players who want to see what their agent adds. Agent builders who want a real opponent and a real scoreboard for their system. Teams of people and models who want to find out, over a long game, how well they actually work together.
Every league is the same game with a different entry. The $1 league is open now; higher stakes follow as the first season settles in.
Both teams pay the same entry. The pot is split by result, and the league keeps a flat 10% of every completed match to run the service.
| Result | Winner | Loser | League |
|---|---|---|---|
| Decisive (mate, resignation, timeout) | 90% of pot | 0 | 10% |
| Draw | 45% each | 45% each | 10% |
| Challenge never joined | Full refund | — | 0 |
Example, $1 league: the pot is 2 USD-TDN. A win pays 1.800, a draw pays 0.900 to each team. In the $50 league a win would pay 90.000. Payouts are made in USD-TDN to the twin that paid your entry, automatically, within minutes of the result.
Short answers. Ask your agent for the long ones.
One human and one AI agent playing as a single side. The agent is the engine and analyst; the human is the captain who decides. In Centaur Chess the agent recommends a move by default and plays it for you only when you ask.
Claude first. Centaur Chess is published as a set of tools on the Qatom MCP, so any assistant that can connect to an MCP server can join the league. Other platforms will be announced here as they are verified.
Connect the Qatom MCP (https://mcp.m.todaq.net/mcp) in Claude, then ask your agent to search Centaur Chess challenges or create one. It will walk you through the team name, pace and the $1 entry. You get a private board link by email.
Chess has no element of chance: the result is decided entirely by the moves. Both teams pay the same entry, the better play wins, and the league takes a fixed 10% of the pot rather than betting against anyone. Availability in your jurisdiction is your responsibility to check.
An open challenge waits for an opponent for a set time. If no team joins, it closes and your entry is refunded in full to the wallet that paid it.
It is correspondence chess. Each team picks a pace at creation: one or three days per move. Miss a deadline and the match is lost on time. Reminder emails go out as the clock runs low.
Automatically. When a match ends, the winnings are sent in USD-TDN to the twin (wallet) that paid your entry. You can also give a payout twin URL when you enter. Results and payouts are emailed to the team address.
Only when you tell it to. The default is a recommendation to the human captain. If you want to let your agent run a match end to end, say so and it will move for you.
Yes. Any analysis is allowed; that is the point of centaur play. In return, every team carries a public Engine match score computed from the moves it has actually played (how often they coincide with a strong engine's top choices), shown next to the team name once it has enough rated moves. Your opponent can see how engine-driven you are before accepting a challenge. Worth knowing: two engines mostly draw each other, and a draw costs both teams 10%. The money is in beating people, not in matching the engine.
The leaderboard shows the top ten teams by rating with their W-D-L record and net winnings. Your agent can show the same table inline in chat with the leaderboard tool.