Gemini’s “agentic trading” lets AI models like ChatGPT and Claude plug into user accounts via MCP, executing crypto trades autonomously and turning AI from signal vendor into primary CEX client.
Summary
- Gemini has wired its full trading API into Anthropic’s Model Context Protocol, so compatible AI agents can pull market data, query order books, place orders and manage positions directly from user‑linked accounts.
- Users set budgets, strategies and caps, while modular “Trading Skills” give agents DCA, grid, multi‑leg and risk tools, making a growing slice of Gemini’s resting and market orders originate from opaque, black‑box models.
- Unlike TON’s non‑custodial “Agentic Wallets,” which push autonomy to Telegram edge wallets, Gemini centralizes agentic activity inside a regulated CEX perimeter, recasting AI as a client type that humans merely configure.
Gemini has rolled out “agentic trading,” a feature that lets AI systems like ChatGPT and Claude connect directly to user accounts and execute crypto trades autonomously on the exchange, rather than just spitting out trade ideas for humans to click. The move quietly shifts AI from being a glorified signal service to being a client class in its own right, with opaque, proprietary models now sourcing, routing, and managing a chunk of CEX order flow on their own.
According to Gemini’s own blog, “agentic trading means your AI agent acts on your behalf — placing trades, monitoring markets, and managing risk automatically,” with users defining strategies and constraints while the agent handles execution. Under the hood, Gemini has integrated its full trading API with the Model Context Protocol (MCP), an open standard originally built by Anthropic that lets AI agents call external tools and services; compatible models include Claude and ChatGPT, which can query markets, place orders and adjust positions over time. Third‑party write‑ups emphasize that Gemini is the first regulated US exchange to expose a dedicated “agentic” interface, turning centralized exchange infrastructure into a native venue for autonomous trading agents rather than just human click‑flow and traditional algos.
Gemini heats up the AI race
Practically, the system is built around modular “Trading Skills” — pre‑built functions AI agents can invoke to get real‑time market data, inspect order‑book depth and spreads, and pull historical candle data, with more complex order‑routing and risk modules promised over time. Users link their accounts to an AI model via MCP, set budget and risk limits, and then let the agent run strategies that can range from simple DCA or grid trading to multi‑leg structures and volatility plays, with Gemini stressing that “human oversight remains part of the design” through caps and rules. But the microstructure implication is obvious: once enough people plug in agents and walk away, a material share of resting and market orders on Gemini will be coming from black‑box models tuned to optimize for particular objectives, not from human decision cycles.
That changes who you are actually trading against. Historically, the story was “retail vs HFT vs a few prop‑shop algos”; now Gemini is effectively advertising “AI as a client type,” more akin to how prime brokers have algorithmic clients that are not directly human‑decisioned on each trade. In high‑volatility periods, tightly coupled agent strategies can amplify feedback loops — especially if many users are copying off the same “AI signals” or fine‑tuning similar models on overlapping data — and you can easily imagine clusters of agents front‑running naive human behavior or unintentionally engaging in coordinated patterns that look a lot like cartelized flow.
There is a clean contrast here with TON’s on‑chain “Agentic Wallets.” TON is pushing autonomy to the network edge: agents live in Telegram, manage non‑custodial wallets on TON, and interact with DeFi directly on an L1. Gemini is doing the opposite: recenters agentic trading inside a regulated, custodial CEX, where AI agents are tightly coupled to one exchange’s API and compliance perimeter. In both cases the future is the same: the next “HFT villain” in crypto will not be a named firm on the other side of your order, but a swarm of un‑audited models, systematically optimized around the fee, tax and KYC constraints their operators face — and increasingly treated by the infrastructure itself as the primary customer, with humans demoted to parameter‑setters and occasional override buttons.
