polybot

Features

Everything polybot does.

A complete picture of the platform — agent integration, strategies, AI models, venues, risk, data, and architecture.

MCP

Agent integration

MCP server (stdio + HTTP)

Native Model Context Protocol server exposing 25+ typed tools. Drop into Claude Desktop, custom agents, or any MCP client.

Read-only introspection

list_markets, get_market, get_positions, get_orders, get_strategy_code — safe defaults so agents can explain what is happening.

Strategy assessment tools

analyze_strategy, compare_strategies, suggest_strategy_improvements — agents can look back 30 days and propose changes.

Approval-gated execution

Three modes — disabled, shadow (paper), live (approval). Every agent-initiated trade is audit-logged.

Human-in-the-loop

Live mode requires explicit approval per order or pre-authorised thresholds. Agents never fill blindly.

10 built-in

Trading strategies

10 battle-tested strategies

Arbitrage, statistical arbitrage, AI model, spread farming, copy trading, resolution arb, calendar spread, momentum, poll divergence, volume spike.

BaseStrategy interface

Subclass, implement on_price_update / on_market_event, return Signals. Scheduler, risk, and execution are handled.

Per-strategy shadow mode

Enable live for arbitrage, keep momentum in paper. Independent state per strategy.

Hot enable / disable

polybot strategy enable <name> / disable <name> — no restart, no downtime.

Pluggable models

AI plugins

AIModelPlugin base class

Return a probability and a confidence. polybot handles signal generation, sizing, and risk gating.

Built-in providers

Anthropic Claude (claude-sonnet-4), OpenAI GPT (gpt-4o), Perplexity (web search + reasoning).

BYO model

Point at any HTTP endpoint or local model. Type-checked input and output; no prompt glue in strategies.

Multi-exchange

Venues

Polymarket CLOB

On-chain USDC, EIP-712 signed orders via py-clob-client. Primary venue.

Kalshi

CFTC-regulated US exchange. Per-event markets.

Opinion

DEX-style execution for long-tail prediction markets.

Binance (hedging)

Spot and perpetuals for delta-hedging prediction positions against correlated crypto exposure.

Invariants

Risk & safety

Position & exposure caps

Per-market, per-category, per-venue, and global USD caps enforced pre-submission.

Daily loss limits

Auto-disable strategy or venue when drawdown threshold is hit.

Shadow-first defaults

New strategies boot in paper mode. Promotion to live is explicit and reversible.

Full audit log

Every signal, every order, every MCP tool call logged to DuckDB with timestamps and actor.

Dual DB

Data & analytics

SQLite operational store

Markets, orders, positions, trades, strategy configs, wallets. ACID transactions.

DuckDB analytics

Price history, trade history, strategy stats, market correlations. OLAP queries in milliseconds.

Prometheus metrics

Strategy P&L, fill rate, latency, error counts — scrape into Grafana.

Async services

Architecture

NNG pub/sub

Scanner, executor, and strategies communicate via nanomsg-next-gen. Decoupled, backpressure-friendly.

FastAPI + WebSocket UI

Vue.js dashboard served from the same Python wheel. Real-time P&L and order flow.

Docker-first

Single docker-compose brings up the whole stack. Runs anywhere Python does.

Need an agent system built like this?

Cryptuon builds production AI agents, MCP integrations, and trading systems. polybot is our open-source showcase.