On 10 July 2026, Revolut connected its standalone crypto exchange, Revolut X, to third-party AI assistants including Claude, Gemini, OpenClaw, and Cursor. Traders can request market analysis, set alerts, backtest strategies, and prepare market or limit orders through natural-language prompts — without living inside the exchange UI.
The important design choice is not the list of models. It is the approval gate: Revolut states that users remain responsible for reviewing and approving every order before execution. The assistant prepares; the human confirms.
This article explains what the integration actually unlocks, how it fits the broader agentic-trading wave, and what builders should copy (and avoid) when wiring AI into financial execution paths.
What Revolut X exposes to assistants
Revolut X is Revolut’s dedicated crypto trading surface — separate from the main banking app — with deeper charts, tighter spreads for active traders, and API access. The AI layer treats that exchange as an execution and data backend, not as a chatbot bolted onto marketing copy.
Reported capabilities through the assistants:
- Portfolio and trade history summaries
- Real-time market data pulls
- Custom price alerts (for example, Telegram when ETH hits a target)
- Market and limit order preparation across 300+ tokens
- Strategy exploration and backtests (for example, a grid strategy on BTC over 30 or 90 days)
Setup is deliberately non-developer-facing for the common path: install a connector, extension, or plugin for the chosen assistant. Users on other platforms can still link via a universal skill or a CLI published on GitHub. Revolut had already shipped AIR, an in-app AI for general financial tasks, in April 2026; this launch is the crypto-trading-specific depth cut.
Think of the assistant as a trading desk junior who can pull blotters, run a backtest, and draft an order ticket — but cannot hit send without a senior signature.
Agentic trading here means conversational preparation of trades, not unsupervised autopilot.
Why the approval gate matters
Revolut’s public language is unusually blunt for a fintech launch: it does not endorse or guarantee third-party AI tools; AI may err; Revolut is not liable for losses, missed opportunities, or erroneous trades caused by those tools. Users must review and approve orders.
That is not just legal boilerplate. It is a product architecture decision:
| Mode | Who decides | Failure mode |
|---|---|---|
| Full autopilot | Model / agent loop | Bad prompt → immediate bad fill |
| Prepare + human approve | Human final authority | Bad prompt → rejected or corrected ticket |
| Assistive analysis only | Human places every order manually | Slow, but no AI on the execution path |
Revolut chose the middle row. That matches how serious agent tooling should meet regulated money movement: models are high-variance; settlement is not. Keeping a human in the loop for execution reduces the blast radius of hallucinated sizes, wrong symbols, or misread risk metrics.
The same pattern is showing up elsewhere. Gemini shipped agentic trading tooling in April 2026; Liquid’s Co-Invest brings live execution into ChatGPT and Claude; Robinhood has previewed crypto-focused agentic accounts after an agentic stock-trading beta. The industry is converging on AI as the front-end, with exchange APIs as the back-end — and liability language that still points at the user.
The stack underneath: skills, connectors, APIs
From a builder’s perspective, Revolut is packaging exchange endpoints behind agent skills — tools the assistant can call under the open Agent Skills pattern — rather than inventing a proprietary chat UI that only works inside Revolut.
That design has consequences:
- The model is swappable. Claude today, Gemini tomorrow, Cursor for people who already live in an IDE-shaped agent.
- The exchange remains the source of truth for balances, open orders, and fills.
- Prompt quality becomes part of trading risk. Ambiguous natural language (“buy some ETH”) is a worse ticket than a structured order form unless the skill forces confirmation of size, side, and type.
For teams building similar products, the durable piece is not “we integrated Claude.” It is a stable skill surface over trading APIs, plus a mandatory confirmation step before any state-changing call.
What this means for builders
If you build fintech or exchange products:
- Prefer prepare → review → execute over autonomous trading loops for retail or semi-pro users.
- Publish skills or MCP-style tools against your existing trading API; do not rebuild the whole product inside the chat transcript.
- Make confirmation UX show symbol, side, size, order type, and estimated fees in machine-checked fields, not only in free-form assistant prose.
- Assume third-party models will be wrong sometimes; write liability and UX for that case up front, as Revolut did.
If you build agent harnesses:
- Treat financial write actions as a separate permission class from read-only market data.
- Log the prompt, tool calls, and human approval artifact for every executed order — you will need them for support and disputes.
Conclusion
Revolut X’s AI assistant integrations are a clear instance of agentic trading done as conversational order preparation, not as unattended bots. Claude, Gemini, OpenClaw, and Cursor become interchangeable front-ends; Revolut X stays the execution venue; the human stays on the approval gate.
That is the pattern worth copying in the stack: open skills over real APIs, model-agnostic connectors, and hard stops before money moves. The models will keep improving. The approval gate should not disappear just because the prose got better.