AI-native trading bot infrastructure

Runbot builds the infrastructure behind AI trading apps. A trader types an idea. Our AI Agent turns it into a strategy the Runbot engine can execute, backtests it on years of tick and order-book data, optimizes it in the background, and deploys it as a bot. Partners launch all of this under their own brand.

• An AI Agent that creates, tests, optimizes and deploys strategies from plain English, with a validation pipeline so that what the model writes is what the engine accepts.
• A strategy engine with 50+ indicators and 200+ trading rules, plus custom rules and indicators that users create themselves, from 1-second to daily timeframes.
• A realistic backtester that walks historical order books for slippage, applies each venue's fee model and checks stops tick by tick.
• Execution on three market types: perpetual futures, short-term Up/Down prediction markets and pari-mutuel contests, without Runbot ever holding user funds.
• A white-label platform: our first partner went live in September 2026.
• A paid MCP API so that other software, including AI agents, can buy strategies per call.

Distribution. Selling one app to one partner reaches a whole community at once. The partner owns the relationship; Runbot owns the engine, the data and the AI.

Margin. The expensive parts (market data, backtesting compute, LLM inference) are built once and shared across every partner and every Strategy Copilot call.

Defensibility. An LLM alone cannot trade. What makes the agent useful is six years of engine grammar, more than 10 TB of market data and a pipeline that turns free text into strategies the engine accepts. That pipeline is the part competitors would have to rebuild.

Timing. Short-term prediction markets have become one of the fastest-growing trading venues in crypto, and they are a natural fit for rules-based bots (chapter 02).

Rules-based bots work, and the engine is the moat. Our first product, HxroBot, automated 1-minute and 5-minute pari-mutuel contracts in 2020 and was acquired by its exchange partner. Runbot then spent 2023–2025 building a backtester that replays tick trades and order books, an optimizer, the Multi-Backtester Tool and a fleet of more than a hundred market-data feed processes. That engine is still the heart of everything in this document.

Combined monthly volume on Polymarket and Kalshi grew from under $5B in September 2025 to about $24B in April 2026¹. In April 2026 alone Polymarket's international venue traded about $9B and its new US venue about $1.3B¹. Capital followed: Intercontinental Exchange, owner of the NYSE, committed up to $2B to Polymarket in October 2025 at a valuation of about $8B², and Kalshi raised $1B at a $22B valuation in 2026³.

Short-term crypto contracts are the bot-shaped part of this market. Polymarket launched 15-minute BTC and ETH "Up or Down" markets in September 2025, settled by Chainlink price feeds⁴, then added other durations. A contract lasting minutes rewards exactly what software does well: reading the underlying market continuously, pricing the outcome, and buying only below a fair price, every window, around the clock. It punishes what humans do: hesitation, fatigue, revenge trades.

US$ billions · source: Pew Research Center / The Block

Crypto derivatives traded about $85.7 trillion in 2025⁵, and derivatives represented roughly 73–77% of centralized-exchange activity in early 2026⁶, almost all of it perpetual futures. On-chain perpetual venues are catching up: Hyperliquid alone reached about 6% of total perpetual volume, including centralized exchanges, in March 2026⁷.

This is where the demand for level-2 bots lives: strategies that combine several indicators, order flow and adaptive risk management, rather than the grid and DCA bots exchanges already give away. Runbot has served this segment since 2023, and it is the segment our history (BTC perpetuals since December 2019) covers deepest.

The Model Context Protocol (MCP), published by Anthropic in November 2024, became the common way for AI assistants to call external tools. OpenAI adopted it in March 2025, and in December 2025 it was donated to the Agentic AI Foundation, a Linux Foundation fund co-founded with Block and OpenAI and backed by Google, Microsoft and AWS⁸. In parallel, Coinbase open-sourced x402 in May 2025, a protocol that lets software pay per HTTP request in stablecoins⁹.

Runbot has offered strategy building to AI agents over x402 on mainnet since July 2026 (chapter 06).

Existing options leave a gap. Exchange-native bots are free but limited to grid, DCA and copy-trading. Quant frameworks require engineers. General-purpose AI chat can write a strategy, but without a quant-grade stack behind it the backtests break, the deployment breaks and the data is not there. Runbot combines a production engine, an AI that speaks its grammar and a white-label delivery model, which is why platforms can launch in weeks rather than build for years.

1. Pew Research Center, "Trading volume on prediction markets has soared", 27 May 2026, based on data from The Block.
2. Intercontinental Exchange, "ICE Announces Strategic Investment in Polymarket", 7 Oct 2025.
3. Kalshi, company announcement of a $1B raise at a $22B valuation, 2026.
4. Polymarket announcement of 15-minute crypto markets, Sept 2025.
5. CoinGlass data via CoinMarketCap Academy, Jan 2026.
6. CCData / CoinDesk Data "Exchange Review", February and March 2026.
7. The Block, "Hyperliquid gains ground on centralized exchanges as perps market share nears 6%", 2026.
8. Linux Foundation and Anthropic announcements, 9 Dec 2025.
9. Coinbase Developer Platform (x402 launch, May 2025).

Since June 2026 the product follows one rule: one description should produce a strategy that is already built, tested and improved. The user does not have to know what a selector or a trailing stop is before seeing a result. The search for better parameters is free; the user pays only to apply a variant they like.

1. 01DescribeAn idea in plain words, or a template
2. 02BuildThe agent writes the strategy's rules one by one
3. 03BacktestA free preview on recent history
4. 04OptimizeA free background search for better variants
5. 05DeployPaper trading or live trading

Build a trading bot by talking to it

• Plain-language creation. "A BTC 5-minute Up/Down strategy that follows momentum and never pays more than 60 cents a share" is a complete request.
• Results in the conversation. Backtests, optimization variants, strategy diffs and confirmations appear as cards inside the chat, each with its price before it runs.
• Operations that survive navigation. Multi-minute backtests and optimizations stay visible across pages and sessions, with notifications when they finish.

• Create a strategy, or start from a template or an uploaded file.
• Explain, edit and compare versions, with a restorable history.
• Run basic and pro backtests, optimizations with a time and speed budget, and multi-strategy backtests with the Multi-Backtester Tool.
• Deploy to paper trading on every market and to live trading where the venue is enabled, then report on running bots from the dashboard.

General-purpose language models are good at describing strategies and bad at producing payloads a trading engine will accept. The Runbot AI Agent closes that gap with structure, validation and memory rather than with a bigger prompt.

1 · Phase-gated conversation

Each conversation moves through phases (understand, design, build, multi-backtest, operate). Each turn loads only the instructions and tools that the current phase and the chosen contract type need, which keeps the model focused and the cost per turn down.

2 · Box-by-box assembly

The model never writes a whole strategy at once. It fills eight small slots of the engine's grammar, each validated on its own, then commits:

3 · Five-stage pre-flight, before any engine call

When the engine still rejects a payload, an isolator bisects the strategy to find the offending element and a small repair model proposes a targeted fix. Automatic corrections are always shown to the user, and a correction is never allowed to remove a component the engine considers valid.

4 · Grounded in strategies that passed

The agent retrieves from a library of tens of thousands of decomposed strategy patterns and a set of golden rules per contract type.

5 · Money-safe by construction

Chat is billed pro rata to the tokens actually used. Before a turn starts, credits are reserved; each model request is capped to what the user's balance can pay for; operations are priced and confirmed before they run; and invariants block optimizations the engine would fail on before the user is charged.

6 · Model-agnostic

The agent uses frontier language models for the conversation and smaller ones for repairs and titles, and can switch provider without changing what the user sees. Models are interchangeable; the grammar, the validation and the corpus are ours, and they encode years of knowledge about how the engine behaves.

A grammar for almost any strategy

Under the conversation sits the Runbot strategy model: 50+ indicators and 200+ trading rules, and custom ones on top. It describes a strategy as connected boxes: what to watch, when to act, how much to risk and how to exit. The AI Agent fills these boxes for the user; power users can still edit them one by one.

Anatomy of a strategy. The scalper built in the conversation shown in 03.1, box by box. Each box is validated on its own before the strategy is saved.

Backtests that respect the order book

A backtest is only useful if it is honest about what a bot would really have paid. The Runbot backtester is event-driven and works at two resolutions at once: 15-second candles drive the indicators, while true tick trades decide whether a stop-loss or take-profit was hit inside a candle.

• Slippage from real depth. A simulated market order walks the historical level-2 order-book snapshot level by level, respecting the venue's minimum size, size increments and maximum size, plus an order delay.
• Every fee model. Maker and taker fees and funding on perpetuals; the probability-shaped taker fee of prediction markets; pool fees on pari-mutuels.
• Settlement like the venue. Up/Down contracts are settled in backtests from Chainlink price data, the oracle source the venue itself relies on.
• The metrics that matter. Compounded APR, monthly performance, Sharpe, profit factor, win rate, risk/reward, max drawdown in % and in value, time in position, variance of PnL and drawdown, the full equity curve and trade list, and the Runbot Score.

Backtest windows depend on plan and contract type (for example up to 5 years on perpetuals).

In September 2026 a 90-day BTC perpetual backtest with EMA crosses, stops and a trailing exit dropped from 35.1 s to 13.3 s. A 180-day pari-mutuel backtest went from 2,089 s to 26 s with byte-identical results. Inside optimization loops, individual backtests run in about a second.

Search the space, keep the robust

The optimizer explores a strategy's parameters with no limit on their number, in search spaces that can exceed 10¹² combinations, with Runbot's in-house optimization.

Objectives are explicit. The user, or the agent on their behalf, weights up to 5 of 12 criteria: win rate, drawdown, flat drawdown, Sharpe, profit factor, Runbot Score, monthly performance, risk/reward, number of trades, time in position, drawdown variance and PnL variance. A minimum-trades floor rejects variants that look good because they barely trade.

The optimization assistant turns a conversation ("optimize my take-profit and stop, one year of history, keep drawdown under 5%") into a search plan sized to the chosen time and speed, then shows two results: the best Runbot Score and the best APR.

In the AI app, a free background auto-optimization follows every new strategy. Paid optimizations are priced per minute of compute, with ×2 and ×4 speed tiers.

The Multi-Backtester Tool backtests up to 20 strategies together, across different markets and pairs, with a capital weight per strategy. It reports each strategy and the combined portfolio side by side, which is how users spot strategies that duplicate each other and ones that hedge each other.

Example: three strategies, one portfolio · illustrative numbers

The trend follower loses in choppy markets, exactly when the mean-reversion strategy earns, so their drawdowns rarely happen at the same time. The Multi-Backtester Tool makes that visible: the portfolio's worst drawdown is smaller than any single strategy's. It shows the opposite case too. Two strategies that are the same bet in disguise move together, and combining them does not reduce the risk.

Runs 24/7. Never holds your funds.

A live bot receives market data in real time, evaluates its strategy exactly as the backtester did, and sends orders through a rail that can trade but cannot withdraw.

From one strategy to a shared library

Raw returns are a poor way to rank strategies: they reward luck and overfitting. The Runbot Score combines monthly performance, win rate, risk/reward, return-to-drawdown, maximum drawdown, trade frequency, time in position and the length of the test into a single 0–100 figure, and a separate live score tracks the period after a strategy was published, so buyers can compare the backtest with what happened afterwards. The library shows where each strategy stands against the community.

The 2025 marketplace is still running on EVM chains in the pro app: open-source and rental strategy NFTs, community indicator packs, royalties, and daily re-backtests of listed strategies. Rental expiry automatically stops the bot that used it.

One library shared by every Runbot-powered app: a strategy published on one partner's app can be discovered, backtested and cloned from another's, with its score and live record attached. Partners get content from day one; creators get reach beyond a single community.

Leveraged contracts without expiry that track an asset's price through funding payments. Bots here are "level 2" strategies: trend following, mean reversion, breakout and scalping systems combining several indicators, order flow and adaptive exits.

• Live execution on Hyperliquid through its native API (03.5), on a history recorded since December 2019.
• Maker/taker fees and funding applied in every backtest; slippage from historical depth.
• Leverage-based sizing, risk-based sizing and 32 exit rules including trailing and break-even logic.

Each contract asks one question: will BTC (or ETH, SOL, XRP) finish the window higher than it started? A share of the winning side pays $1, the losing side $0, so the price of a share is the market's implied probability. Runbot supports Polymarket windows of 5 minutes, 15 minutes, 1 hour, 4 hours and 1 day.

• Signals on the underlying. Indicators run on the continuous perpetual feed and are projected onto the contract window.
• Price-aware entries. Rules cap the price paid per share, so a bot trades only where its model sees positive expected value.

Pool-based BTC/USD Over/Under contests of 20 and 60 seconds on Solana, now on glgl.fun, powered by OddsFi. Runbot first automated this pool format in 2020 with HxroBot. Winners share the pool.

The partner collects everything its users pay. Runbot invoices the partner monthly, at wholesale, for what the app consumes:

• AI usage, which the partner resells with its own margin.
• Engine compute: backtests and optimizations, per run.
• A platform fee that covers operations and support.

Pricing follows usage, so a partner can start small and grow with its community.

1. 01DiscoveryAudience, markets, revenue model and jurisdictions.
2. 02AgreementScope, pricing and the markets to open.
3. 03BrandUpload the logo, set the name, point the domain.
4. 04Prices & paymentsSet user prices and connect the partner's own payment accounts.
5. 05Go liveRunbot deploys and verifies the app. The partner markets it.

Most white-label software forces a choice between running everything for the partner and handing over the source code. Runbot does neither. The partner owns its users, its data and its revenue; the intelligence that makes the product work stays with Runbot, and Runbot operates the whole platform.

The partner owns

• Its users, their accounts and their data
• Its brand, its domain and its app
• Its prices, its payment accounts and its revenue
• Its community and its marketing
• Its offering towards its users, and the compliance that goes with it

Runbot keeps and runs

• The AI Agent, its rules and its strategy corpus
• The strategy engine, backtester and optimizer
• Market data and execution rails
• Maintenance, monitoring, upgrades and scaling
• Never handed over, neither as source nor as a build

Each partner is isolated from the others: its own brand, users, payment accounts and data. The partner needs no developers and never operates the platform. Its job is its community.

An agent can hand Runbot a whole task in one prompt, or keep the reasoning and call the tools one by one: build, backtest, optimize and export.

• "Find a BTC strategy with an APR of at least 40% and a drawdown under 15%." A complete task: Runbot builds, backtests, optimizes and exports.
• "Build a Polymarket 5-minute Up/Down strategy that never pays more than 0.60 a share." Another complete task, on a prediction market.
• Build, then optimize for APR ≥ 40%, then backtest the best variant over 365 days, then export. The same goal as tool calls, with the agent choosing each step.

# the user's own agent plans, then calls each tool
→ build      "ETH mean reversion on Hyperliquid, 15-minute candles"
← strategy ready · first backtest: APR 18.2% · drawdown 11.4%
→ optimize   goal: APR ≥ 40% · drawdown ≤ 15%
← 2 variants · best APR 41.7% · drawdown 13.9%
→ backtest   best variant · last 365 days
← APR 38.9% · drawdown 14.2%
→ export
← import-ready strategy

A paid agent service at mcp.runbot.io. Any AI agent can describe a strategy and receive a built, backtested, optimized and exportable result, paying per call over x402 in USDT on X Layer or USDC on Base. Settled on mainnet since July 2026.

# 1 · ask; the 402 reply carries the payment challenge
POST https://mcp.runbot.io/v1/quick-strategy
{ "description": "BTC 5-minute Up/Down on Polymarket,
   momentum, never pay more than 0.60 a share",
  "backtest_window_days": 30 }
← 402 Payment Required · $3.99 · x402

# 2 · replay with a signed payment header
← 200 { "job_id": "…" }   settled only on success

# 3 · poll for free until completed
GET  https://mcp.runbot.io/v1/quick-strategy/{job_id}
← baseline vs optimized metrics + export

Designed for software buyers

• Settle on success only. A failed or clarification-needed call is not charged; one payment buys one build.
• Sandboxed buyers. Each paying wallet maps to its own isolated service account with its own strategies.
• No internals exposed. Buyers receive metrics, a summary and an export, never the agent's prompts or catalogs.
• REST first, MCP compatible. REST is canonical; an MCP JSON-RPC surface serves MCP-native clients.

A backtest can only be as honest as the data it replays. Runbot records its own history rather than buying candles: level-2 order books, every trade, candles from 15 seconds to one day, funding rates, open interest, liquidations and prediction-market order books.

The history spans bull and bear markets alike. Every regime sent something back into the engine: an execution edge case, a data gap, a way a strategy breaks that nobody predicts on a whiteboard. Feeds are watched continuously, and a stalled feed alerts the team.

Most retail tools fill a backtest at the candle close, at any size, instantly. Runbot prices every instant from the last real trade and fills market orders against the order book that actually existed.

Average fill 104.26, not 104.20. Six cents. On a bot that trades thousands of times a month, those cents are the strategy's edge, kept or destroyed. Illustrative book.

• Funds stay with the user. On Hyperliquid, on the user's own account. On Polymarket, in a wallet the user owns and funds from their own external wallet. On pari-mutuel venues, on the user's venue account.
• Trade, never withdraw. Hyperliquid agent wallets cannot withdraw by construction, and Polymarket orders are signed by a delegated key whose policy forbids withdrawals.
• Keys encrypted at rest. Venue credentials are encrypted per user and redacted from logs.
• Per-user engine identities. Each user's live bots run under their own engine profile, isolated from other users and other partners.
• Consent before money moves. Prices are shown and confirmed before paid operations run, automatic corrections are disclosed, and nothing goes live without an explicit deploy by the user. Deleting an account stops its live bots first.

• Strict isolation of each user's data, verified by automated tests; every money operation goes through audited functions.
• Personal-data retention rules; users can delete their account.
• Jurisdiction-based access restrictions on partner apps and in production.
• Replicated data and rehearsed restores.

Runbot provides software. It does not hold client assets, does not execute on its own account for clients and does not give investment advice. Users decide what to deploy and trade on their own accounts. Partners are responsible for the compliance of their own offering in the jurisdictions they serve, and partner apps apply geographic restrictions.

A partner runs the app as its own business: it sets the prices, bills its users as merchant of record and keeps the revenue.

Runbot takes no fee on the volume that partners' users trade.

Each partner brings a community, each community creates strategies, and each tested strategy makes the agent and the library better for every other partner. Better results keep users trading, and partners earn more on their volume.

1. 01Partners launch appsand bring their communities
2. 02Users build & runstrategies, backtests, optimizations, bots
3. 03Revenue and datapartner revenue, Copilot calls, strategies
4. 04Better agent, richerlibrary, better results → retention for every partner

Held, not just locked

Only 20.6% of the supply circulates outside the vesting contract. About a third more is unlocked and simply not withdrawn: the airdrop pool sits undistributed, and so do much of the incentives and treasury pockets.

Release rates are read from the vesting contract, whose balance reconciles exactly with the sum of its 91 vesting schedules. September 2026.

The 2025 staking tiers belonged to a product that no longer exists and have been retired. Locking stays in production on-chain, and the new utility ties the token to how the platform is used today.

Phase 1 · The licence gate

Phase 2 · Native, own brand

Phase 1 works with the business as it is today: partners license the platform, so partners are where demand for the token starts. Phase 2 lets end users pay for credits in RBOT directly, at a discount.

Supply, vesting and locks are all on-chain. Contract addresses are listed in appendix B. Any change to the allocation, the vesting schedule or the utility model will be announced publicly before it takes effect.

$RBOT is not a share and confers no ownership, profit, dividend, voting or governance right, and no claim against Runbot SAS. The utility model above is a design intention, subject to announcement before launch; it can change.

1. 2020

   HxroBot, a no-code bot builder for 1-minute and 5-minute pari-mutuel contracts, later acquired by its exchange partner.

2. 2023 – 2026

   Runbot engine: backtester and optimizer, then level-2 slippage, tick-level event loop and the NFT rental marketplace on Arbitrum.

3. 2025

   $RBOT TGE on Arbitrum One; staking and vesting contracts live.

4. 2025

   AI optimization assistant; multi-market engine for perpetuals, Polymarket and pari-mutuels.

5. 2026

   Engine v2 with live event streams, the base of every Runbot-powered app.

6. 2026

   AI Agent moves to engine v2; "one-shot" direction; Polymarket live execution.

7. 2026

   Strategy Copilot live on mainnet over x402; per-token chat billing.

8. 2026

   Hyperliquid native execution live; first white-label partner live.

Alexandre Papazoglou · Founder

Serial entrepreneur and crypto enthusiast, with a Master's in Economics and Digital Strategies. Ran operations at HxroBot, a trading-automation product acquired by its exchange partner, then founded Runbot, where he now leads strategy, partnerships and product. x.com/SeimanRunbot

The team that built HxroBot in 2020 went on to design the Runbot engine and its NFT marketplace, and has run trading automation on real capital ever since.

Engineering. The team maintains and extends the Runbot engine, the AI layer, market data and execution rails.

Nature of this document

This white paper is published by Runbot SAS for information purposes only. It describes the company's products, technology and plans as of September 2026. It is not an offer to sell or a solicitation to buy any financial instrument or crypto-asset and is not a prospectus.

No advice

Nothing in this document is investment, financial, legal or tax advice. Runbot provides software; users decide what to build, deploy and trade, on their own accounts and at their own risk.

Trading risks

Crypto-assets, perpetual futures and prediction markets are highly volatile and can lead to the partial or total loss of the capital committed, and leverage amplifies losses. Backtests and past performance do not guarantee future results; live results can differ from backtests because of liquidity, latency, fees, venue outages or model error. AI-generated strategies can be wrong. Screens marked as demo data and figures marked as illustrative are not real results.

Forward-looking statements

Roadmap items, planned features and planned token utilities are intentions, not commitments. They may change, be delayed or be abandoned because of market, technical, partner or regulatory developments.

$RBOT

The $RBOT token contract is 0x53ce5F2A18A623317357d6366eC59b47cbb0060D, on Arbitrum One. $RBOT is a utility token. It confers no equity, ownership, profit, dividend, voting or governance right and no claim against Runbot SAS or any affiliate. Its market value is not guaranteed and can fall to zero. Token utilities are operational benefits that Runbot may modify, suspend or replace; the utility model in chapter 10 is planned and subject to legal review. Smart contracts can contain vulnerabilities; blockchain transactions are irreversible; users are solely responsible for their wallets and keys. Staking is governed by the Terms and Conditions of the staking interface.

Availability

Runbot products and partner apps are not offered in jurisdictions where they would be unlawful, and access may be restricted by geography. Partner apps are operated independently by their owners, who are responsible for their own offering and compliance.

Third parties

Third-party names, trademarks and data belong to their owners and are used for identification only. Market figures from third-party sources are cited as published and have not been audited by Runbot.

Publisher

Runbot SAS, société par actions simplifiée registered with the Rennes Trade and Companies Register under number 922 047 238, registered office 90 bis rue de Fougères, Rennes, France. Director of publication: Alexandre Papazoglou. Contact: contact@runbot.io.

Runbot White Paper, 2026 edition, version 2.0, September 2026. Supersedes the 2025 edition.

Always verify a contract address against runbot.io before interacting with it. Runbot will never ask for your seed phrase or private key.