Your AI Trading Agent Sounds Great. The Wallet Still Needs You.
Wallets still need hands. That's the part of the agentic DeFi pitch the press release glosses over.
1inch Business, the developer-facing arm of the decentralized exchange aggregator serving 27 million users and processing over $300 million in daily trades, announced on March 30 an expansion of its Model Context Protocol server. The MCP server, first launched last month as a way to connect AI coding assistants to the 1inch ecosystem, now gives trading agents direct access to the full suite of 15 APIs: swap execution, portfolio data, gas costs, token metadata, and order books included.
The Model Context Protocol, developed by Anthropic and now an open standard, has become the connective tissue for agent-to-tool communication across a widening range of domains. In 1inch's implementation, an agent can query liquidity across decentralized exchange pools, build a swap execution plan, and retrieve code examples for integrating the API, all through natural language. The server exposes seven tools: search, list_examples, get_example, swap, orderbook, spot_price, and product_api. Four require authentication. Three are public. The setup works with Cursor, VS Code, Claude, Codex, JetBrains, and Gemini out of the box.
"This is real infrastructure," is what the architecture says. "Agents will be executing the majority of swaps by 2030," is what the 1inch co-founder says. These are not the same sentence.
Sergej Kunz, 1inch co-founder, told the company's blog: "Agents, not humans, will be executing the majority of swaps by 2030." He followed with a qualifier the headline omits: "The agent economy cannot eliminate market competition. Poorly informed agents will underperform skilled humans. That is why choosing the infrastructure around the agent is as important as the strategy."
That's a defensible position dressed as a prediction. The more instructive sentence is the one that follows it: the 1inch MCP server is non-custodial and does not execute transactions on behalf of any party. Read the architecture docs and this becomes concrete. The swap tool builds execution steps and returns a quote. The orderbook tool constructs a limit order server-side. Both require the developer to sign the returned typed data and submit it back through the orderbook endpoint. The agent plans. The human key signs. The blockchain settles. Kunz himself is the citation for the 2030 number; the architecture is the citation for everything the machine can and cannot do today.
This is not a criticism of the product. The MCP server is a genuine step forward for agent-accessible DeFi infrastructure. Giving agents structured access to on-chain data, pool liquidity, gas costs, cross-chain routing, is a real engineering problem, and 1inch has shipped working code with broad tool support. The GitHub repository hosts the MCP server config and the Agent Skills package, which packages setup instructions so an agent can connect without copy-pasting documentation. The hosted MCP endpoint is at https://api.1inch.com/mcp/protocol.
The gap between "agents will execute most swaps by 2030" and "the MCP server is non-custodial and does not execute transactions" is not a contradiction. It's a dependency graph. More agent-accessible infrastructure means more agents can analyze, plan, and optimize. Fewer humans need to click buttons. But the button still exists, and someone with a private key still has to press it. That's not an agent infra problem: that's how ECDSA signatures work, the cryptographic algorithm securing Ethereum and most DeFi transactions. The press release language tends to blur this; the architecture docs do not.
What changes as MCP becomes the standard layer for agent-to-DeFi communication is the developer experience for the humans building agentic trading workflows. Before the protocol, a developer integrating an AI assistant with on-chain exchange data had to write custom API wrappers, manage authentication flows, and handle the impedance mismatch between natural language planning and signed blockchain transactions. Now the agent speaks MCP natively, and the developer writes the signing layer. The integration work that precedes execution gets faster: days to minutes, as 1inch puts it.
Whether that maps to "agents executing the majority of swaps" by any specific year is a separate question, and one the infrastructure alone cannot answer. It depends on what happens to private key management, whether hardware wallets adopt agent-native signing workflows, and whether regulators treat agent-initiated transactions the same as human-initiated ones. These are open problems. 1inch has solved the API access layer. The key management layer is a different story, owned by a different set of standards bodies and wallet developers.
For now, the practical use case for the 1inch MCP is the one the architecture actually supports: agents that can see, plan, and recommend, attached to human operators who still hold the keys. That's genuinely useful: portfolio monitoring agents that surface arbitrage opportunities, trading assistants that draft execution plans for human approval, multi-agent coordination across chain types. The non-custodial constraint is not a limitation of this specific product; it is a structural feature of blockchain UX that agent frameworks have not broken, and probably should not try to.
The MCP standard is accumulating adoption across domains, and 1inch's DeFi implementation is the latest example. The protocol has graduated from coding assistant connectors to live financial infrastructure: worth noting as a data point in that trajectory, separate from the vendor predictions attached to any single launch. The predictions are what the company wants you to print. The architecture is what they shipped.