Workshop · 45–60 minutes

Games on Markets.

A workshop on building games where a live prediction market is the game engine: odds as game state, hourly markets as game loops, and trading agents as players. Runs live at builder events and works self-serve: everything on this page can be built today, starting with no API key.

Self-serve friendly · No API key needed to start

Skip to what you can build →

The idea in one breath

01

A prediction market is socially writable game state. Reality supplies the content, traders update the state, expiry supplies the clock, and resolution supplies the judge. The order book publishes what the crowd believes right now, moves in real time as people trade, and settles to an objective outcome. Read one way, that’s a trading screen. Read another, it’s a number between 0 and 1 that a crowd with money on the line is fighting over, and your game can read it for free.

You also get opponents without scripting them: the traders are already playing against your game. This workshop is a tour of what that unlocks, games you can build on Limitless market data in an afternoon with an AI agent doing the typing, and where to go when you want real stakes in the loop.

What the market gives your game

03

A live probability feed

Market data on Limitless is public: prices and order books over REST and WebSocket, no API key required to read. Your game can consume real crowd-priced odds the way it would consume a physics engine.

A clock

Crypto price markets recur on hourly and daily cadences, so a market is always about to resolve: a natural round timer. Esports and football markets settle shortly after the match, fast enough for a session-length game loop.

Real stakes, when you want them

When the game calls for skin in the game, the full trading API is there: signed orders, four SDKs (TypeScript, Python, Go, Rust), and account history. Everything is documented at docs.limitless.exchange.

People keep building this

05

Every mechanic in this workshop already works somewhere. Click through these before you design anything, then notice what none of them can do:

Streaks · Underdog

Pick eleven in a row, win up to 1,000× your entry (the advertised cap has moved between 1,000× and 5,000× depending on format). Underdog has since launched its own regulated prediction-market exchange. Play one entry to feel why a streak counter is hard to walk away from.
underdogsports.com/games/streaks

Prize leaderboards · Esports World Cup

The 2026 Esports World Cup put $300,000 into free-to-play pick’em and fantasy leaderboards. No entry fee, no stake of any kind. Demand for esports prediction games is not the open question; what these players cannot do is hold a position.
esportsworldcup.com

Jackpots · Megapot

A $1-a-ticket daily jackpot on Base paid in USDC, with 19 winners in its first year and a biggest payout around $200,000. It is permissionless and ships a forkable kit, so a jackpot can sit on top of whatever you build.
megapot.io · a campaign built on it

Frontends · Azuro

By mid-2024, 27 third-party frontends were running on Azuro, five of them past $100,000 in revenue. Building consumer surfaces on shared market infrastructure is a working business. The structural difference: Azuro pools liquidity and prices the odds, so those frontends are bookmakers rather than order books.
azuro.org

And on Limitless · Viral.games

Players bet on which social posts would go viral, with real markets underneath, aggregating Polymarket and Limitless. An ETHOnline 2024 finalist, quiet since. It was built before the order book, the recurring short-cadence markets, and the SDKs existed, which is most of what makes this workshop possible.
the ETHGlobal showcase entry

Streaks, prize leaderboards, jackpots, third-party frontends: all of it works. We have not yet seen it on an order book, where players hold real positions and the market settles in minutes.

What you can build

03 · tiers

The order book hands you more than one number, and all of them are public. Probability is the obvious one; how fast it moves, how much the two sides disagree, how much money is behind each, and how long until the answer are all readable too. What any of them becomes in your game is the actual design work. The last column is only there to get you unstuck:

Signal What it tells you One thing it could drive
ProbabilityHow confident the crowd is, right nowAnything that should feel safe or precarious
Price velocityHow fast that confidence is movingMomentum, panic, escalation
SpreadHow much the two sides disagreeUnreliable readings, risk, noise
Order-book depthHow much money sits behind each sideWeight: how hard something is to shift
Time to expiryHow long until you find outRound length and the pressure curve
ResolutionThe answer, settled objectivelyConsequences that stick

The bar for market-native: the design should break if you swap the live market for a random number generator. If it survives the swap, it’s a fine first build, not proof of the thesis.

0

Read the market

No key · an afternoon

Public market data only. No wallet, no funds, nothing to custody. Each of these is small enough to finish in a sitting, with an agent doing most of the typing.

Win-probability overlay
A second screen or stream widget showing live odds for an esports or football match while it plays, the way a broadcast shows win probability.
The market is the boss
Game state driven by live odds: loot tables, weather, or enemy waves that scale with the hourly BTC market. Nobody has to author the difficulty curve.
Pick’em arcade
Players lock picks on real markets and score by the payout multiplier at resolution: the cheaper the side you correctly picked, the more points. Contrarian skill wins. No funds move.
Survivor
Pick right on consecutive markets or you are out. Last player standing wins, and the market cadence sets the round length for you.
1

Know the player

Player brings an address

The player enters just their public wallet address and the game reads their public trading record: volume, positions, accuracy. Their real conviction becomes the character sheet. Never ask players to paste API keys into a game; anything authenticated belongs server-side.

Prediction pet
A tamagotchi that thrives when your predictions resolve right and sulks when they don’t.
Accuracy RPG
XP from resolved positions, with character classes by market category: esports main, macro main, hourly degen.
2

Trade from code

SDKs · real orders

Programs that place real orders through the SDKs. You do not start from an empty file: agents-starter is the public reference implementation, with the signing, order placement and run loop already solved, and its AGENTS.md and QUICKSTART.md are the current source of truth. This tier involves real money and real risk: start dry, size small, and read the risk material in Agents Academy before going live.

Agent arena
Everyone forks agents-starter, swaps in their own strategy, and the bots run head-to-head on the same market set with fixed bankrolls, ranked on risk-adjusted results. The plumbing is shared, so the game is strategy design: signals, sizing, execution, risk.
Auto-hedged game show
A trivia or skill game whose host lays off outcome risk on the live market, so the house survives a hot streak. Market making as game design.

Mechanics that travel well

02

Two design patterns that pair with almost any prediction game, and map neatly onto the tracks hackathons and game jams tend to run:

Sealed picks

Prediction games have an obvious use for onchain confidentiality: commit picks confidentially and reveal at resolution, so nobody can copy the leader mid-round. Blind pick’em pools, hidden-role prediction games, and secret-conviction leaderboards all fall out of that one trick.

Jackpot layers

Jackpot mechanisms compose naturally with prediction accuracy: streaks of correct picks as the entry mechanic, a jackpot as the payout curve’s fat tail. Skill decides who enters the draw; the draw keeps casual players in the room.

The live build

~20 min

The workshop’s centerpiece is a Tier 0 build done live: an AI agent one-shots a pick’em app against public market data, the room enters picks on an hourly market, and the market resolves on screen before Q&A ends. The same prompt works self-serve. Paste it into your agent harness of choice (Claude Code, Cursor, anything that can read docs and write files):

Paste into your agent
Read the Limitless developer docs at https://docs.limitless.exchange. Using only public market data (no API key, no wallet, no orders), build a single-page pick'em app: list Limitless's recurring up-or-down crypto markets (they run on 5-minute, 15-minute, and hourly cadences), let a player enter a display name and pick UP or DOWN on the market closest to resolving, save picks in localStorage, and when the market resolves, score each correct pick 1 divided by the price of its side at entry time, capped at 20, wrong picks score zero, and show a leaderboard. Vanilla HTML and JavaScript, one page, no build step. One catch: the market data is public but not CORS-open to outside browser origins, so serve the page from a tiny local server that proxies the API reads.

Round two, because iteration is the actual skill: turn the pick’em into a bot arena. Three house strategies playing every round against the room is a Tier 0 agent arena, no keys, no orders. The Tier 2 version of the same idea trades real orders through the SDKs.

Then iterate
Now add three house bots that play every round automatically: MoMo Bot always picks the crowd favorite (the side priced above 50 percent), Fade Bot always picks the underdog, Coinflip Bot flips. Bots enter their picks a couple of seconds after each new market loads, show up in the picks list and on the leaderboard like any player, and can be toggled on and off.

The first prompt sends your agent to docs.limitless.exchange on purpose: the docs are canonical and evolve with the API, so the agent reads the current truth instead of this page hard-coding endpoints that go stale.

Play the reference build: both prompts applied, running against live markets right now on 5-minute, 15-minute, hourly, and daily windows, bots included.