Share
The hardest part of decentralized finance (DeFi) has never been the finance. It has been the interface. Swapping one token for another requires connecting a wallet, navigating a decentralized exchange, approving a spending limit, confirming the swap transaction, waiting for block confirmation, and repeating the process if anything fails.
Doing this across multiple protocols, several chains, and a portfolio of positions requires either deep technical familiarity or expensive help. For the roughly 100 million people globally who hold crypto but do not consider themselves power users, that complexity is the reason they have never touched DeFi at all.
Base MCP is Coinbase’s answer to that barrier. Launched on May 26, 2026, it connects Base smart wallets to AI agents including ChatGPT, Claude, and Cursor through the Model Context Protocol, turning conversational interfaces into execution layers for onchain transactions. The user types what they want in English and the agent proposes the transaction. The user approves or rejects it through the Base app. Nothing executes without that final human confirmation.
The Model Context Protocol, known as MCP, is an open standard that Anthropic introduced in late 2024 to solve a specific problem in AI infrastructure. Large language models (LLMs) are powerful at reasoning but blind to live data and incapable of taking actions in external systems without a defined connection. MCP is that connection: a universal protocol that lets AI models plug into external tools, APIs, databases, and applications through a standardized interface.
Before MCP, every integration between an AI model and an external tool required custom code written specifically for that combination. A developer building a ChatGPT integration with a DeFi protocol needed to write, maintain, and secure a bespoke connection for each protocol, each action type, and each AI model. MCP replaces all of that with a single standard. Any MCP-compatible AI client can connect to any MCP server without custom integration work for each pair.
Base MCP is Coinbase’s implementation of the Model Context Protocol, exposing onchain actions to any MCP compatible AI agent. Rather than being limited to ChatGPT or Claude, it works with any AI client that supports the standard, giving the broader MCP ecosystem blockchain access through a single integration. Coinbase designed Base MCP as foundational infrastructure rather than a standalone product feature.
The most important design decision in Base MCP is the one most users will not notice: it never touches your private keys.
Authentication between an AI agent and a user’s Base Account uses OAuth 2.1, the same authentication standard that governs secure connections between third-party apps and services like Google or GitHub. When a user connects their Base Account to an MCP-compatible AI agent, they authorize the connection through their account, not by handing over any cryptographic secret. The MCP server holds no private keys and cannot sign or broadcast transactions on its own authority.
When an agent proposes a transaction, it generates a link that opens the Base App. The user sees a simulated preview of the exact asset changes the transaction will produce before any confirmation is required. They can review the output, check that the numbers match what they intended, and then approve or cancel the action. The transaction only hits the blockchain after that explicit approval. The AI agent is a proposal engine, not an autonomous executor.
This architecture addresses the specific security risks that earlier wallet-connected AI experiments introduced. Prompt injection attacks, where malicious content in a webpage or document tricks an AI into proposing an unintended transaction, are mitigated because every sensitive action requires a manual approval step that the user must complete in the Base App rather than inside the AI interface.
Phishing and domain hijacking risks are reduced because transactions are proposed within the AI framework rather than on external websites where spoofed interfaces can mislead users.
Base MCP launched with six DeFi protocol integrations, each delivered as a modular skill plugin that tells a compatible AI agent how to call that protocol’s specific functions.
Developers can extend the system by writing custom skill plugins using markdown specifications. The spec tells the agent how to call an external API and how to hand off the resulting unsigned transaction to the user for approval. Any DeFi protocol, onchain application, or API that can be described in a markdown specification can become part of a user’s Base MCP environment without waiting for Coinbase to build the integration.
The clearest illustration of what Base MCP enables is a workflow that previously required a user to visit two separate protocol interfaces, check current supply rates on each, calculate the difference, withdraw from one, and deposit into the other while managing approvals at every step.
With Base MCP, a user opens ChatGPT, connects their Base Account through the MCP connection, and asks: “Which protocol is paying higher USDC yield right now, Morpho or Moonwell, and move my funds to whichever one pays more.”
The agent queries current supply rates from both protocols through their respective skill plugins. It finds that Morpho is currently offering a higher annualised rate on USDC. It proposes a sequence of two transactions: first, withdraw the user’s current Moonwell USDC deposit; second, deposit the resulting funds into Morpho’s USDC market. The agent presents both transactions with their expected asset changes before either executes.
The user reviews the proposal, confirms both steps in the Base App, and the two transactions execute onchain. What previously required fifteen minutes of navigation across two protocol interfaces, two sets of transaction approvals, and enough familiarity with lending market mechanics to make the comparison meaningfully, takes under two minutes and requires no protocol knowledge beyond the plain-language instruction.
Coinbase’s Head of AI Product Lincoln Murr emphasized that Base accounts sync transaction history and portfolio data whether users are in the AI agent interface or the Base app, meaning the conversational interface and the native application share the same state. There is no reconciliation problem, no duplicate position tracking, and no delay between an onchain action and its reflection in the user’s portfolio view.
— lincoln.base.eth (@MurrLincoln) June 13, 2026
Base MCP launched into a market where AI agent activity is already measurable at production scale. A May 2026 report co-published by Keyrock found that AI agents settled approximately 176 million onchain transactions totalling more than $73 million between May 2025 and April 2026. The typical transaction was a few dozen cents, and approximately 98.6% were denominated in USDC, a distribution that reflects both the stablecoin’s dominance in agent payment flows and the micro-settlement character of most agent-to-agent commerce.
Coinbase’s Base network alone processed more than 119 million x402 protocol payments across 480,000 active agents by June 2026. The x402 integration inside Base MCP means agents can pay for x402-gated APIs and services directly from a connected wallet within the same conversational session, extending the interface from user-facing DeFi actions into the autonomous agent payment layer simultaneously.
MCP 2026-07-28 is live and it's the largest update to the protocol since launch.
MCP is now stateless, making it easier to deploy and scale remote servers. https://t.co/K8KqxbUh4e
— ClaudeDevs (@ClaudeDevs) July 28, 2026
The security risks that accompany that scale are also already visible. In the same period Base MCP launched, Bankr, one of the six launch partners, halted token transfers after 14 wallets were compromised, an incident that illustrated how prompt injection attacks and overbroad approval flows represent the next wave of DeFi security concerns in agentic environments.
Every Base MCP transaction requires explicit user approval before execution, a design choice that directly addresses the approval-layer vulnerability the Bankr incident exposed.
Base MCP is not an autonomous trading system. The human approval requirement means it cannot execute strategies independently, rebalance positions overnight, or respond to price movements without the user actively confirming each transaction.
For users who want fully autonomous agent execution, Coinbase’s AgentKit and Agentic Wallet products built on the same infrastructure are designed for that use case, with programmable spending limits that allow pre-authorised actions within defined parameters.
Base MCP is also not chain-agnostic. It operates on Base, Coinbase’s Ethereum layer-2 network. Balance and transaction history lookups support other EVM-compatible chains, but transaction execution is limited to the Base network at launch. Cross-chain execution through Chainlink CCIP or other bridge infrastructure is not natively supported through the MCP interface.
And Base MCP is not a self-contained discovery layer. The agent knows how to use the six launch protocols and any custom skill plugins a developer adds. It does not independently browse the DeFi ecosystem, evaluate new protocols, or identify yield opportunities outside the protocols its skill plugins cover. The quality of what it can do is bounded by the quality of the skill specifications it has been given.
Those limits are features rather than failures of ambition. A system that requires human approval at every sensitive step and operates within a defined protocol set is a system that can be trusted at the point where DeFi has historically been least trustworthy: the moment a new user decides whether to connect their wallet to something they do not fully understand.
Share
preferred on
