CloddsBot's README contains one line that explains the whole project: Claude + Odds = Clodds. The name captures the essence — use Claude as the brain, use odds (probabilistic markets) as the battlefield, and build a conversational trading terminal.
12 days of development, 10.7k clones in 14 days, 377 stars in a single day. It was built for the Colosseum Agent Hackathon (a Solana hackathon), but its polish doesn't look like a 12-day product.
21 Chat Platforms × 10 Prediction Markets × 7 Futures Exchanges
CloddsBot's coverage is somewhat absurd:
- Messaging channels: 21 platforms including Telegram, Discord, WhatsApp, Slack, Teams, Signal, Matrix, iMessage, LINE, and Nostr, plus a built-in WebChat
- Prediction markets: Polymarket, Kalshi, Betfair, Smarkets, Drift, Manifold, Metaculus, PredictIt, Opinion.xyz, Predict.fun
- Perpetual futures: Binance, Bybit, Hyperliquid, MEXC, Drift, Percolator, Lighter — up to 200x leverage
- On-chain: the full Solana stack (Jupiter, Pump.fun, Raydium, Orca, Bags.fm) plus five EVM chains (Base, Ethereum, Arbitrum, Optimism, Polygon) via Uniswap V3, 1inch, and Virtuals Protocol
- Bittensor subnet mining: TAO
- Momentum, mean reversion, penny clipper, expiry fade strategies
- DCA bots, smart routing, whale tracking, copy trading
- Arbitrage detection, cross-market hedging
- 5-minute, 15-minute, 1-hour, 4-hour, and daily round strategies for Polymarket BTC/ETH/SOL/XRP binary markets
- Circuit breakers: daily loss limits trigger automatic shutdown
- VaR/CVaR: value-at-risk and conditional VaR to quantify tail risk
- Volatility regime detection: separating normal volatility from anomalies
- Stress testing: replay of historical extreme scenarios
- Kelly sizing: mathematically derived position sizing based on win rate and odds
- Kill switch: one-click liquidation of all positions
- Honeypot detection (buyable but not sellable)
- Rug-pull analysis (withdrawable liquidity)
- Holder concentration (whales can dump)
- Risk scoring
This isn't a trading bot — it's a trading operating system. You issue commands from any chat platform, Claude understands the intent, invokes one or more of 121 skills, and executes the trade.
121 Skills: Trading Strategies as Composable Blocks
The most interesting design in CloddsBot is its skill system. 121 built-in skills, each an independent trading capability unit:
This design is isomorphic to Claude Code's skill system — don't build one big monolithic agent; build many small, focused skills and let the agent call them on demand. The benefits: composability, extensibility, and independent testability.
But CloddsBot's skills aren't written for developers — they're written for trading strategies themselves. Each skill encapsulates a complete trading logic (entry conditions, exit conditions, position management, risk control). Claude's job is to understand user intent, pick the right skill, and pass the right parameters.
Risk Engine: Core, Not Optional
The biggest risk with a trading agent isn't losing money — it's losing money without knowing it. CloddsBot's risk engine does several things:
In a traditional quant firm, all of this is the job of an entire risk department. CloddsBot compresses it into an open-source project anyone can run.
Token Safety: Rug-Pull Defense
On-chain trading carries a risk that doesn't exist in traditional finance — the token itself may be a scam. CloddsBot integrates GoPlus for on-chain token audits:
MCP Server Mode: Trading Directly from Claude Desktop
CloddsBot can run as an MCP server; clodds mcp install configures Claude Desktop / Claude Code in one step. This means you can say "buy BTC up on Polymarket" inside Claude Desktop, and Claude calls CloddsBot's MCP tools to execute the trade.
This is the MCP protocol landing in finance. MCP isn't just for reading files and querying databases — it lets Claude place orders on your behalf. Of course, that also means: if your Claude gets phished, your money goes with it.
What 12 Days Can Build
CloddsBot was built for the Solana hackathon in 12 days. But the author clearly didn't start from zero — 121 skills, 21 messaging channels, 10 prediction markets far exceed 12 days of integration work. The more likely explanation: the author was already building trading bots, and the hackathon was the occasion to consolidate existing components into an open-source product.
That also explains why the project gained 377 stars in a day — it's not a demo, it's a usable production-grade tool.
What's Worth Watching
CloddsBot isn't the first trading bot, but it gets several things right:
1. Skill system instead of monolithic agent — 121 independent skills that compose and extend, isomorphic to Claude Code 2. Chat as trading — any of 21 platforms, lowering the barrier to use 3. Risk engine as core, not add-on — VaR/CVaR/Kelly/circuit breakers, quant-firm-grade risk control 4. MCP server mode — Claude Desktop trades directly; MCP landing in finance 5. Token safety auditing — "automated due diligence" for on-chain trading
The risk is equally obvious: for a Claude-driven trading agent, the biggest risk is Claude itself. If prompt injection makes Claude execute unintended operations, or Claude's reasoning fails under extreme market conditions, the losses are real money. CloddsBot's risk engine addresses market risk, not AI safety risk.
But as an open-source project, it proves one thing: the combination of MCP + a skill system is enough to build a trading terminal covering 1,000+ markets. The extensibility of this architecture goes far beyond what traditional trading bots imagined.
---
GitHub: https://github.com/alsk1992/CloddsBot