Retail
Bybit ships an AI agent built into the exchange itself, betting against the MCP-connector pattern every rival just adopted
Bybit AI, launched 9 September 2026, trades spot, futures and options and staffs customer support from inside the app, running in a ring-fenced sub-account rather than connecting an outside model like Claude or ChatGPT the way Binance, Coinbase, Kraken and OKX now do.
The crypto-exchange story this year has been one of near-total convergence: Binance's Agent OS, Coinbase's agent, Gemini's Agentic Trading and Kraken and OKX's own connectors all opened the door the same way, letting a customer point an external model such as Claude or ChatGPT at their account through the Model Context Protocol. Bybit AI, launched 9 September, breaks that pattern deliberately. CTO Rockman Zhang's framing, that the goal is a single interface rather than "another dashboard" or a "bolt on chatbot", reads as a direct rejection of the MCP approach the rest of the field settled on, in favour of an agent Bybit builds, hosts and controls end to end.
That is the same fork in the road this site flagged when PYMNTS surveyed Scalable Capital, Robinhood, Interactive Brokers and Public on equities: whether the reasoning layer between account data and a live trade sits with the platform or with a general-purpose model the customer already uses for everything else. Bybit's answer is closer to Robinhood's Agentic Trading than to Questrade's MCP connector or Coinbase's model, in the concrete sense that both use a ring-fenced sub-account with its own budget to cap blast radius rather than routing every action through a confirmation click or an outside model's judgment. CEO Ben Zhou has already flagged a version 2 that adds external connectors and model choice, which suggests the built-in launch is a starting position, not a permanent rejection of the industry's chosen architecture.
The trade-off is the same one this site keeps returning to whichever asset class is involved. A platform-built, platform-hosted agent is bounded entirely by whatever guardrails and data access Bybit itself chose to give it, with no visibility into a third-party model's training or incentives, at the cost of ceding the entire reasoning layer to the exchange rather than to a model the customer picked and already trusts elsewhere. An MCP-connected agent inverts that: the customer keeps the model, but the exchange has less control over what a general-purpose assistant might do with account access once granted. Neither Bybit nor its MCP-based rivals have published incident data on live agent trades, so for now the sub-account-budget design and the external-connector design are both live experiments running on real customer money, distinguished mainly by who is holding the reasoning layer when something goes wrong.
Read the original: Cryptowisser - Bybit ships an AI agent built into the exchange itself, betting against the MCP-connector pattern every rival just adopted. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.