CloddsBot: An Open-Source AI Trading Agent

CloddsBot is an open-source AI agent that trades. Not "analyses" or "assists": the README's own description is an agent that "scans for edge, executes instantly, manages risk while you sleep", pointed at prediction markets, crypto spot, perpetual futures and Solana token launches. It is self-hosted, MIT licensed, built on Claude, and it crossed 1,745 GitHub stars by 11 September 2026 after a v1.9.0 release at the end of August.

This is a report on what CloddsBot is and what it asks you to trust. It is not a recommendation, and there is nothing here about configuring it to trade. Nothing below is financial advice.

Key Takeaways:

• CloddsBot is a self-hosted terminal that places real orders, not a signal feed.
• It was built in 12 days for a hackathon, and the repo is eight months old.
• The risk controls it advertises are self-declared, and nobody outside the project has audited them.
• MIT licensing means the software comes with an explicit disclaimer of warranty, which matters more here than in most repos.

What CloddsBot is

The name is a contraction the project explains itself: Claude plus odds. What you install is an npm package called clodds that runs as a local gateway, with a browser interface at localhost:18789/webchat and connections out to messaging platforms, so you can talk to it from Telegram or Discord rather than a terminal. Node 22 or newer, TypeScript, SQLite for local state, LanceDB for semantic memory, PostgreSQL for analytics.

One thing to check before you run the install command, because it surprised me: the newest release tagged on GitHub is v1.9.0, dated 31 August 2026, but the latest version published to the npm registry is 1.7.7. The one-line install the README leads with therefore hands you a build from February, not the one the release notes describe. Whether that gap in CloddsBot's release process is deliberate or a publish step nobody noticed, I cannot tell you; either way, npm install -g clodds and git clone are not currently getting you the same software.

The project describes itself on cloddsbot.com as "the intelligence layer for odds markets", and the homepage carries a pair of buttons reading "I'm a human" and "I'm an agent", which tells you something about the intended audience. There is also an agent-only forum where, in the project's description, AI agents discuss markets and vote on content while humans can read but not register.

The repository was created on 26 January 2026 and had commits landing hours before this was written, so it is actively worked on: 447 commits, 14 tagged releases, 21 open issues and 9 open pull requests. Seven months old and busy. Both halves of that matter later.

Which markets it reaches, and how the numbers add up

The headline claim is 1,000-plus markets. That number is not itemised anywhere I could find, and the itemised lists are smaller and more useful: 10 prediction market venues named individually as Polymarket, Kalshi, Betfair, Smarkets, Drift, Manifold, Metaculus, PredictIt, Opinion.xyz and Predict.fun, plus 7 perpetual futures exchanges including Binance, Bybit, Hyperliquid and MEXC, plus Solana and EVM decentralised exchanges. Reading the two together, "1,000+ markets" appears to count individual contracts rather than venues, which is a fair thing to count and a confusing thing to put on a badge next to numbers that count venues.

While checking those figures I found a small inconsistency worth naming: the README badge says 119+ skills, the feature table says 119 bundled, and cloddsbot.com's stat strip says 118. One off by one, across the project's own surfaces.

That is a tiny discrepancy and it would be unfair to make much of it, except for one thing: the whole proposition here is that you hand this software exchange credentials and go to sleep. The standard of care you want from a project in that position is higher than the standard you want from a CLI that renames files. Noticing whether a project's numbers agree with each other is a cheap proxy for how carefully it is maintained, and it is the sort of check we apply to any agent component before wiring it into something that matters, the same instinct behind reviewing what third-party agent skills are permitted to do.

How the agent is put together

Underneath, CloddsBot is a skills-shaped agent rather than a monolithic trading bot, which is the genuinely interesting part of its architecture. The project bundles roughly 119 skills, lazily loaded, and exposes all of them through an MCP server so Claude Desktop or Claude Code can drive them directly. It claims 8 LLM providers, 4 specialised agents, and 18 tools.

That design decision is why the feature list reads the way it does. Whale tracking across six chains, copy trading with sizing controls, backtesting with stop-loss and take-profit validation, token security audits through GoPlus for honeypot and holder-concentration checks, a code-scanning shield with 75 rules and a scam address database: none of those are trading logic, they are capabilities a general agent can call. If you have read our piece on what agent skills are, CloddsBot is one of the more aggressive demonstrations of the pattern in the wild, which is what makes it worth reporting on even if you would never run it.

The always-on requirement is not negotiable in a design like this. Prediction markets on Polymarket settle in 5-minute, 15-minute, hourly and daily rounds, and an agent that misses the round has no function. That means a machine that stays awake, which a laptop is not, and it is the one requirement the README quietly assumes and never addresses. Zentor is a hosted cloud AI computer built precisely for the assumption that goes unstated there: an agent process that keeps running when your own machine does not, on infrastructure separate from your daily driver.

The risk surface, stated plainly

Start with what the project says it does about risk, then with what it cannot do anything about.

The README advertises a unified risk engine: a circuit breaker, VaR and CVaR calculations, volatility regime detection, stress testing, Kelly sizing, daily loss limits and a kill switch. That is a serious-sounding list and it is exactly the list a serious implementation would have. It is also entirely self-declared. No third party has published an audit of whether those controls fire correctly under load, and the honest position for a reader is that the list describes intent rather than verified behaviour.

Four things sit outside what any risk engine can cover. First, self-hosting means you hold the keys: exchange API credentials and wallet keys live on your machine, and the blast radius of that box being compromised is your funds rather than your data. Second, MIT licensing includes the standard clause that the software is provided as is, without warranty of any kind, and that the authors are not liable for any claim or damages. People skim that line in every repo; in a repo that moves money it is the most load-bearing sentence in the file. Third, the perpetual futures integration advertises up to 200x leverage, which is a capability, not a recommendation, and it does not become safer because an agent is the one using it. Fourth, this was built in 12 days for the Colosseum Agent Hackathon, by the project's own account, and seven months of subsequent commits does not convert a sprint into a battle-tested system.

The first of those four is the one you can actually do something about without writing any code, and it is worth separating out. Credential blast radius is a placement question, not a code-quality question: the same software is a different risk on a machine that also holds your password manager, your email session and your SSH keys than it is on a box that holds nothing else. Giving an agent its own computer is the cheapest control available here, and it is the shape Zentor takes as a product, a hosted machine the agent owns rather than a process running as you on the laptop you carry around.

One detail cuts the other way and deserves credit: the arbitrage module defaults to dry-run mode. A default that does nothing until you opt in is the right default, and it is rarer than it should be.

What the arbitrage claim is actually based on

The arbitrage detection is documented as an implementation of arXiv paper 2508.03474, and that citation checks out: the paper is "Unravelling the Probabilistic Forest: Arbitrage in Prediction Markets". Citing a real paper for a real technique is not nothing, given how many trading repos cite nothing at all.

The idea it implements is simple enough to state. If the YES contract on an event trades at 45 cents and the NO contract at 52, buying both costs 97 cents to guarantee a dollar, and the three-cent gap is the edge. Across venues the same logic applies when Polymarket prices an event at 52 and Kalshi prices it at 55. The README is upfront that cross-platform arbitrage carries currency and settlement complexity, which is the polite way of saying the three cents can evaporate between the two legs.

Who this is for, and who it is not for

CloddsBot is for developers, and that is a narrower audience than the marketing implies. You install it with npm, you hold your own keys, you read the trading docs, and when something breaks at 2am you are the operator. The chat-from-Telegram interface makes it feel like a consumer product; the failure modes are not consumer failure modes. If you would not be comfortable reading the source of the module that sizes a position, this is not software for you.

If you do want to look closer, read in this order: the AUDIT.md and security policy the repo ships, then the trading documentation on execution and safety controls, then the risk engine source itself. CloddsBot makes those easy to find, which is one more thing in its favour, and reading them is how you turn a claim about a kill switch into knowledge about a kill switch.

For everyone else, there is a more defensible version of the same interest, which is to separate the research from the execution. An agent that watches markets, gathers the FedWatch and polling data CloddsBot pulls in, and writes you a summary every morning is a genuinely useful thing that risks nothing. An agent that acts on that summary without you is a different decision, and it should be made separately and deliberately. Running the first half on a Zentor cloud computer keeps the always-on monitoring going without putting exchange credentials on a machine that also holds your email, and the boundary between reading and acting stays where you put it. We looked at a research-first take on the same territory in our write-up of TradingAgents.

FAQ

Is CloddsBot free and open source?

Yes. The LICENSE file is the MIT License, copyright 2026 alsk1992, which includes the usual disclaimer that the software carries no warranty and the authors accept no liability.

Does it trade automatically with real funds?

That is what it is built to do. The project describes autonomous execution across prediction markets, futures and decentralised exchanges, with risk controls applied by its own engine. The arbitrage module specifically defaults to dry-run. Treat any statement about automated execution as the project's description of its software as of September 2026, and read the trading docs yourself before it touches an account.

Which exchanges does it connect to?

Named prediction market venues include Polymarket, Kalshi, Betfair, Smarkets, Drift, Manifold, Metaculus, PredictIt, Opinion.xyz and Predict.fun. Perpetual futures venues include Binance, Bybit, Hyperliquid, MEXC, Drift, Percolator and Lighter. Solana and EVM decentralised exchange integrations are listed separately.

Which version do I actually get from npm?

As of 11 September 2026, clodds on npm is at 1.7.7 while the GitHub repository's newest tag is v1.9.0 from 31 August. If you want what the recent release notes describe, build from source rather than installing the package.

Do I need a Claude API key?

The onboarding wizard asks for one, and the project is built on Claude, though it lists support for 8 LLM providers.

Is this safe to run?

Nobody outside the project can answer that for you, and the honest summary is that it is seven-month-old software holding exchange credentials, shipped under a licence that disclaims all warranty. The dry-run defaults and the kill switch are good signs; they are not an audit.