AI-Trader: an agent-native trading platform where AI agents register themselves

What AI-Trader actually is, and the traders it is built for

Most trading tools assume a human at the keyboard. AI-Trader inverts that. The README opens with the claim that "AI agents need their own" trading platform, and the product is built around that premise: an agent gets an account, publishes signals and strategies, joins discussions, copies other agents' positions and earns points for predictions that work out.

There are two front doors. A human signs up at ai4trade.ai with an email and browses signals or follows performers. An agent joins by being told to read a URL. That second path is the whole design thesis. The README states that any agent joins "in seconds" by receiving the message that points it at https://ai4trade.ai/SKILL.md, and it lists OpenClaw, nanobot, Claude Code, Codex and Cursor as supported agents.

The audience is therefore narrow but real: people who already have an agent running somewhere and want it to do something with market data, plus traders who want to mirror agents rather than other humans. If you have no agent and no interest in building one, the human side is a signal feed with paper trading attached, and the README does not describe anything there that a dozen other platforms lack.

The self-registration mechanism: a URL the agent is told to read

The onboarding is the most distinctive thing in the repository. An agent is handed a single instruction, which the README prints verbatim, and the README says the agent will then read the integration guide, install the necessary components and register itself on the platform.

That is a skill-file pattern. The repository backs it with a skills/ directory holding several SKILL.md files: ai4trade for the main integration, copytrade for the follower side, tradesync for the provider side. There is also docs/README_AGENT.md for integration and docs/api/openapi.yaml as the full API specification. So the agent is not guessing. It is reading a document the maintainers wrote for it, then calling documented endpoints.

The architecture in the README is small: skills/ for agent skill definitions, docs/api/ for OpenAPI specs, service/ split into server/ (FastAPI backend) and frontend/ (React), and assets/ for images. The environment file confirms the shape of the runtime. It sets CLAWTRADER_CORS_ORIGINS to http://localhost:3000 and https://ai4trade.ai, which tells you the frontend is expected on port 3000 and the hosted service is the production origin. Background jobs are configurable: POSITION_REFRESH_INTERVAL=300, POLYMARKET_SETTLE_INTERVAL=60, MARKET_NEWS_REFRESH_INTERVAL=900, MACRO_SIGNAL_REFRESH_INTERVAL=900, ETF_FLOW_REFRESH_INTERVAL=900, with MAX_PARALLEL_PRICE_FETCH=5 capping concurrent price calls.

The April 2026 release notes describe the FastAPI web service being split from background workers so that user-facing pages and health checks stay responsive while prices, profit history, settlements and market-intel jobs run out of band. That is a sensible split, and the interval values above are the knobs you inherit from it.

Installing AI-Trader and registering an agent for the first time

The README does not give a pip install line or a Dockerfile. What it does give is the self-hosting path for the database, and the agent path for onboarding. Start with the database, because the backend will not run without a decision there.

The README says to copy .env.example to .env and choose exactly one backend. If DATABASE_URL is set, PostgreSQL is used and DB_PATH is ignored. If it is empty, SQLite is used at the path in DB_PATH. The example file spells out the precedence in a comment and gives a sample connection string.

cp .env.example .env
# PostgreSQL: set DATABASE_URL and leave DB_PATH unused
# DATABASE_URL=postgresql://ai_trader:[email protected]:5432/ai_trader
# SQLite: leave DATABASE_URL empty, the API uses DB_PATH
# DB_PATH=service/server/data/clawtrader.db

Market data needs a key. The example file ships ALPHA_VANTAGE_API_KEY=demo, which is a placeholder rather than a working key. ADANOS_API_KEY is optional and described as enriching US stock market-intel snapshots with social, news and prediction-market sentiment. The June 2026 release notes add that AI-Trader prefers Alpha Vantage but falls back to yfinance when that key is missing, rate-limited, or returns no usable price. That fallback is the difference between a dashboard with numbers and a dashboard with gaps.

For the agent side, the README gives the exact instruction to hand your agent. The same message includes a compatibility alias.

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

What you should see, according to the README, is the agent reading the guide, installing components and registering itself. It then gains the ability to publish signals, join discussions, copy trades, sync signals across brokers and earn points. If your agent cannot fetch a URL or follow a multi-step instruction, none of this happens, and you are back to the human signup flow.

Paper trading, Polymarket, and where the automation stops

The README is careful about one word: paper. New traders are offered $100K paper trading with simulated capital. Polymarket support, added in March 2026, is described as paper trading with real market data and simulated execution, with auto-settlement handling resolved markets through background processing.

Read that together with the broker list. The README says AI-Trader is compatible with Binance, Coinbase and Interactive Brokers, and one feature is described as keeping your broker, syncing your trades and sharing signals. Sync is not execution. Nothing in the README describes AI-Trader placing an order on your behalf at a broker. The signal types are Strategies for discussion, Operations for copying and Discussions for collaboration, which are publications, not order tickets.

So the honest description of the automation boundary is this: agents generate and share signals, the platform scores them with live mark-to-market pricing, and copy trading mirrors positions. If you want an agent that sends a real order to a real broker, this repository does not document that path, and the README's own framing of copy trading as following top performers and mirroring their positions leaves the execution venue unstated.

That is not a flaw in a research platform. It is a flaw if you bought the headline. The title says fully automated trading. The body describes a signal exchange with simulated fills.

Experiments, challenges and the scoring machinery

The recent release notes are mostly about measurement, which tells you where the maintainers spend their time. June 2026 brought improved experiment and challenge progress tracking: expired active experiments now auto-complete on experiment reads, monthly challenges can be created with MONTHLY_CHALLENGE_EXPERIMENT_KEY, and the Experiment Console shows linked challenge performance by variant using the same live mark-to-market scoring as leaderboards.

May 2026 added experiment notice exposure tracking, so prompts shown to agents can be measured separately from explicit message reads. That distinction matters if you are running an experiment and want to know whether an agent ignored a notice or never saw it. It is a research-instrumentation feature, not a trading feature, and it is the clearest signal that this project is as much a study of agent behaviour as a trading venue.

May 2026 also describes a capacity and worker-throttling upgrade for the live service, and April 2026 describes production stability hardening. Both point at a hosted service under real load. If you self-host, MAX_PARALLEL_PRICE_FETCH=5 is the throttle you have to reason about yourself.

What the repository does not settle

Three things are unresolved, and each one affects an adoption decision.

First, licensing. The README carries a licence badge linking to LICENSE, but the top-level repository listing contains no LICENSE file. The project metadata lists the licence as unknown. A badge is not a licence file, and until the file exists in the tree, treat the terms as undetermined.

Second, releases. The repository has no published GitHub releases, so there is no versioned artifact to pin. The last push was on 2026-06-11, and the newest note in the README is the same 2026-06-11 entry. That is not abandonment, but it is also not a project shipping weekly.

Third, operational surface. The example environment file is long, with intervals for news, macro signals, ETF flows and stock analysis. Each of those is a background job that can fail quietly. The README does not document rollback, does not describe what happens to open positions when a worker dies mid-settlement, and does not say how to reconcile the SQLite path against a PostgreSQL deployment if you migrate. Those are the questions to ask before running this against anything you care about.

Alternatives, and the difference that matters

The obvious comparison is a conventional algorithmic trading framework such as Freqtrade or a backtesting library like Backtrader. The difference is not language, it is the unit of work.

A framework like Freqtrade takes a strategy you wrote, runs it against exchange APIs and places orders. The artifact is a bot with a strategy file. AI-Trader takes an agent you already have, gives it an account and a signal feed, and scores what it says. The artifact is a track record attached to an agent identity. Freqtrade answers "did my strategy make money." AI-Trader answers "which agent's calls were worth following," and it answers that with paper capital.

A second comparison is a social trading platform with copy trading, where humans publish and humans follow. AI-Trader keeps the social graph but makes the publishers agents. Whether that is better depends entirely on whether you believe an agent's published reasoning is more inspectable than a human's. The README's bet is that it is, and the skills/ directory is the evidence: the behaviour is written down in SKILL.md files rather than hidden in someone's head.

Editorial conclusion

Adopt AI-Trader if you already run an agent framework and want a place for it to publish signals, debate strategies and mirror other agents' positions against paper capital. Skip it if you expect an execution engine that routes real orders to your broker: the README documents signal sync, not order placement, and the repository carries no LICENSE file even though the README badge links to one. Before pointing an agent at the live platform, read skills/ai4trade/SKILL.md and docs/api/openapi.yaml, then stand up the FastAPI service locally with DATABASE_URL pointing at PostgreSQL so you can see exactly which endpoints the agent calls and what it writes.

Frequently asked questions

Official sources

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