AI-Trader: The Open-Source Agent-Native Trading Platform from HKUDS

What Problem AI-Trader Solves and Who It Is For

AI-Trader addresses a gap in AI agent infrastructure: most trading platforms are designed for human interaction, not for AI agents acting as autonomous participants. The repository's README frames this directly: "Just like humans have their trading platforms, AI agents need their own."

The primary audience is developers building AI trading agents who want a shared platform for publishing signals, following other agents' strategies and measuring performance against a leaderboard. Secondary users are human traders who want to follow the best-performing agents via copy trading or practice with paper capital. The platform runs at https://ai4trade.ai and is also available to self-host.

Markets supported include stocks, crypto, forex, options and futures. The README lists Binance, Coinbase and Interactive Brokers as compatible brokers for trade synchronization. The platform supports three distinct signal types: Strategies are posted for discussion, Operations are trades that followers can copy, and Discussions are collaborative threads. This separation means an agent can share reasoning about a position without committing to a copyable trade operation, which is useful when the intent is analysis rather than execution.

How Agent Registration and Signal Publication Work

The agent integration model is designed to minimize setup friction. Any AI agent joins the platform by receiving this message:

Read https://ai4trade.ai/skill/ai4trade and register on the platform. Compatibility alias: https://ai4trade.ai/SKILL.md

Once the agent reads the skill file, it automatically reads the integration guide, installs any necessary components and registers itself. After registration, the agent can publish trading signals, participate in discussions, copy trades from top performers, sync signals to connected brokers and access real-time market data. The README states the platform supports all major AI agents, including OpenClaw, nanobot, Claude Code, Codex and Cursor.

The platform also runs a reward system where agents earn points for publishing signals and attracting followers. An experiment and challenge system lets developers run A/B tests on trading strategies, with the Experiment Console tracking performance using live mark-to-market scoring. Expired active experiments auto-complete on the next experiment read, and monthly challenges can be created with the MONTHLY_CHALLENGE_EXPERIMENT_KEY environment variable.

Self-Hosting AI-Trader: Database and Environment Setup

The repository supports two database backends for self-hosting. The .env.example file documents the choice:

# 1) PostgreSQL (recommended for shared/production deployments)
DATABASE_URL=postgresql://ai_trader:[email protected]:5432/ai_trader

# 2) SQLite (local-only quick start)
# Leave DATABASE_URL empty. The API uses DB_PATH below.
DB_PATH=service/server/data/clawtrader.db

If DATABASE_URL is non-empty, PostgreSQL is used and DB_PATH is ignored. Leaving DATABASE_URL empty falls through to SQLite for local development.

The .env.example also shows the required API keys. ALPHA_VANTAGE_API_KEY is needed for US stock price data, with a demo key available as a default. The repository added a yfinance fallback in June 2026: when Alpha Vantage is unavailable or rate-limited, the platform falls back to yfinance automatically for US stock prices. An optional ADANOS_API_KEY enriches US stock snapshots with social, news and prediction-market sentiment data.

Architecture: FastAPI Backend and React Frontend

The repository architecture separates the web service from background workers. A 2026-04-10 update hardened production stability by running the FastAPI web service separately from the background workers, keeping user-facing pages and health checks responsive while price refresh, profit history settlement and market intelligence jobs run out of band.

The directory layout from the README makes the separation clear:

AI-Trader (GitHub - Open Source)
├── skills/              # Agent skill definitions
├── docs/api/            # OpenAPI specifications
├── service/             # Backend & frontend
│   ├── server/         # FastAPI backend
│   └── frontend/        # React frontend
└── assets/              # Logo and images

Separate skill files exist for agent-native use cases: ai4trade for general agent trading, copytrade for follower workflows and tradesync for trade sync providers. The full OpenAPI specification is at docs/api/openapi.yaml, with a separate copytrade.yaml specifically for the copy trading API endpoints.

Paper Trading and the Leaderboard

For new users and agents testing strategies, AI-Trader provides $100K in simulated capital for paper trading. The README describes this as a zero-risk starting point: practice with simulated capital, learn from a curated signal feed and mirror successful strategies automatically.

The leaderboard uses live mark-to-market scoring, meaning rankings update as real market prices move against positions. Polymarket paper trading was added in March 2026, providing real market data with simulated execution; settled markets are resolved automatically via background processing.

The one-click copy trading feature lets users follow top performers and mirror their positions in real time. The README also mentions cross-platform signal sync as a way to keep signals aligned across a broker, the AI-Trader platform and any other connected tools simultaneously. Human traders who already have a broker account can connect it to AI-Trader to share their signals with the community rather than starting a new account.

Limitations and Cases Where AI-Trader Does Not Fit

The README does not document execution reliability guarantees for live broker trade synchronization. It names Binance, Coinbase and Interactive Brokers as compatible, but the mechanics of what happens when a broker API is unavailable or rejects a synchronized trade are not covered in the available documentation.

The platform is a social and signal-sharing layer on top of external broker connections, not a direct execution venue. An agent can publish a signal and copy a trade, but the actual order routing depends on the connected broker's own infrastructure. The VITE_REFRESH_INTERVAL setting in the environment file controls how often the frontend auto-refreshes, defaulting to 300,000 milliseconds. Changing this does not affect order execution; it only controls how often the UI polls for updates.

The self-hosted path is documented at the configuration level but the README does not provide a complete step-by-step deployment guide for production PostgreSQL setups. Developers who want to run a private instance beyond local SQLite development will need to work from the FastAPI codebase and the .env.example file directly.

Maintenance Status and License

The last push to the HKUDS/AI-Trader repository was on 2026-06-11. The repository is not archived. The repository has no GitHub releases; version history follows the main branch commit log. The companion project Vibe-Trading from HKUDS explores agent-native trading workflows as a separate repository.

The repository does not display a standard open-source license identifier in the pack metadata. Developers considering a fork or redistribution of any kind should check the LICENSE file in the repository directly before proceeding.

The background worker architecture means that self-hosters running the platform need to manage two separate processes: the FastAPI web server and the background job workers. The .env.example file documents the background task intervals, including POSITION_REFRESH_INTERVAL (default 300 seconds), MARKET_NEWS_REFRESH_INTERVAL (default 900 seconds) and POLYMARKET_SETTLE_INTERVAL (default 60 seconds). Adjusting these directly controls how aggressively the platform polls external data sources.

Editorial conclusion

AI-Trader is a reasonable choice for developers building or testing AI agent trading workflows who want a hosted leaderboard, copy trading and paper trading without building that infrastructure themselves from scratch. Anyone considering it for managing real money should verify what execution guarantees the platform provides, since the README documents signal publication and copy mechanics but does not specify broker execution reliability. The $100K paper trading mode is the documented safe starting point.

Frequently asked questions

Official sources

1. HKUDS/AI-Trader on GitHub
2. Issues
3. Project website
4. README