AI agents are coming for your DeFi trades. Sergej Kunz, co-founder of 1inch, says they'll be doing the majority of swaps by 2030. He's probably underestimating.
On March 27, 2026, 1inch opened its full Business Portal API suite to AI agents via the Model Context Protocol — 15 endpoints including swap execution, orderbook access, portfolio tracking, and gas price estimation. The MCP server is live at api.1inch.com/mcp/protocol, confirmed via the MIT-licensed repo on GitHub with commits as recent as March 29. This isn't a demo or a landing page with a waitlist. The plumbing exists.
The protocol exposes 6 tools to MCP clients: three public — search, list_examples, get_example — and three gated behind OAuth: swap, orderbook, and business_api. The distinction matters. The swap tool is OAuth-gated, meaning agents need authorized credentials before routing a transaction. This is not a fire-and-forget bot framework. It's production-grade access control for a system that moves real capital.
For context on what's actually available: 1inch's Business Portal offers 15 discrete APIs covering Swap, Orderbook, History, Traces, Portfolio, Balance, Gas Price, Token, Spot Price, Transaction, NFT, Web3 RPC, Domain, Token Details, and Charts. The MCP server exposes a subset of those as tools. The 15-API figure and the 6-tools figure describe different things — one is the breadth of the API suite, the other is what's currently callable via MCP. Readers should know which they're tracking.
The non-custodial framing is worth sitting with. 1inch's MCP server does not execute transactions on behalf of any party. It constructs, signs, and submits them through the agent's own wallet. This is a meaningful contrast to Coinbase's AWAC or Trust Wallet's TWAK, which involve varying degrees of custodial or TEE-backed agent custody. In those systems, the exchange or wallet provider holds the keys and acts on the agent's behalf. In 1inch's model, the agent is a sophisticated user with direct blockchain access. Whether that's more or less secure depends entirely on how the agent's key management is implemented — and that's a question the DeFi agent stack is only beginning to answer.
The MCP ecosystem context matters here. Anthropic open-sourced the protocol on November 25, 2024; OpenAI adopted it in March 2025. As of March 2026, the SDK is running 97 million monthly downloads. BitGo launched its own MCP server on March 23, 2026, and reports over 10,000 active MCP servers across public and enterprise deployments. This is not early-adopter territory. The protocol has infrastructure credibility and an installed base. 1inch joining it is a signal, not an experiment.
On the business side: 27 million users and over $300 million in daily trades, per Dune Analytics — third-party data, widely cited, worth noting isn't official exchange figures. The protocol supports 14 chains: Ethereum, Solana, Base, Binance, zkSync, Gnosis, Optimism, Polygon, Linea, Sonic, Unichain, Arbitrum, Avalanche, and Monad. That's a wide surface area for agentic execution across both EVM-compatible and non-EVM chains.
What does agent-driven swap execution actually change? A few things. First, latency. Agents can monitor gas prices, arbitrage spreads, and liquidity shifts in real time — decisions that take humans seconds or minutes happen in milliseconds. Second, MEV dynamics. If agents are executing at machine speed with perfect information, the arbitrage landscape shifts. Liquidity providers and protocols designed for human trading rhythms may find themselves front-run by agents that never sleep and never panic-sell. Third, the capital question Kunz has flagged: "80-90% of capital sitting in liquidity pools is not actually working". Agent-driven swap execution could be the mechanism that forces that capital to actually work — or it could accelerate liquidity sloshing in ways that make pools more fragile. Both are plausible.
Aqua Protocol, which launched November 17, 2025, is attempting a shared liquidity model that speaks directly to this — making liquidity accessible to agents as a first-class concern rather than an afterthought. If agentic execution becomes dominant, how liquidity is structured and who captures value from it changes fundamentally.
The counterargument is straightforward: we've heard this timeline before. 2025 was supposed to be the year of mass AI agent adoption in crypto. It wasn't. Wallets are still confusing, key management is still hard, and the average DeFi user is a human with a Coinbase account. 2030 is a long way out. Regulatory uncertainty, smart contract risk, and the genuine difficulty of building reliable agentic systems all argue for skepticism. The MCP server existing is not the same as millions of agents actually running swaps through it.
That said, the infrastructure being built here is real. The distinction between "read" and "write" operations crossing the agent boundary — from data queries to actual execution — is a genuine threshold. Coinbase's AWAC, Trust Wallet's TWAK, and now 1inch's MCP are forming the primitives of an agent-custody stack. Whether that stack holds under adversarial conditions at scale is an open question. But the foundations are being laid faster than the narrative suggests.
Kunz's 2030 prediction may be early. But the direction is not in doubt.
