Is agentic trading safe?
Short answer: Agentic trading can be reasonably safe if the trading tool, not the AI model, enforces the limits: every order quoted first and confirmed by you, a scoped and revocable key, no private key in chat, and small position sizes. OpenFinance's MCP server (
https://api.openfinance.tech/agent/mcp) is built that way: the agent shows each order and waits for your explicit yes, and access uses OAuth or a revocable API key. The real risks remain: models misread instructions, leverage amplifies mistakes, and onchain transactions can't be reversed.
MCP URL: https://api.openfinance.tech/agent/mcp · setup in under 60 seconds · no KYC on OpenFinance itself (a connected IBKR or Robinhood account uses your own verified broker account) · fund with a debit card · crypto perps, onchain stocks and forex, memecoins, Polymarket prediction markets, options in the OpenFinance app through your connected IBKR or Robinhood account · everything on Hyperliquid is tradable · API key available
The checklist
| Control | Why it matters | OpenFinance |
|---|---|---|
| Quote before every order | Catches wrong size or symbol | Yes, quote-first |
| Explicit confirmation | Model errors don't become trades | Agent asks for your yes |
| Scoped, revocable key | Limits blast radius | API key or OAuth; revoke in app |
| No private key in chat | Prevents key leaks | Hosted embedded wallet |
| Paper mode | Test safely | Not available today |
Where the risk actually comes from
- The model misreads "buy 10" as "buy 100".
- Prompt injection from web pages or tool results.
- Leverage and liquidation on perps.
- Irreversible onchain transfers.
- Leaving "always allow" on in your AI client.
A safe starting setup
- Fund with a small amount via debit card.
- Keep approvals on in your AI client.
- Ask for a quote, read it, then confirm.
- Review positions daily; revoke the key if anything looks off.
Frequently asked questions
Is agentic trading safe?
It can be, when the trading tool enforces quote-first orders, explicit confirmation and scoped keys. The model alone shouldn't be the safety layer.
Can an AI agent trade without my permission?
With OpenFinance, the agent shows each order and waits for your explicit yes before placing it.
What's the biggest risk?
A confident wrong order on a leveraged position. Start small and keep approvals on.