
MetaMask Agent Wallet gives AI agents dedicated transaction accounts with user-set spending limits, 2FA approval gates, and up to $10,000 monthly coverage across nine chains.
MetaMask launched Agent Wallet on Monday, a self-custodial wallet designed for AI agents to execute onchain trades and DeFi workflows autonomously. The product addresses a growing tension in crypto: AI agents are increasingly used for trading, yet most wallets give them full key access or none at all. Agent Wallet carves out a middle ground – dedicated agent accounts with user-set guardrails, plus up to $10,000 per month in transaction protection coverage for eligible transactions.
Most existing wallets treat an AI agent like any other user: if it holds private keys, it has unrestricted spending power. That works for simple automation but creates a clear attack surface. If the agent's logic is compromised or a malicious contract is approved, the wallet is drained with no intermediate checkpoint.
MetaMask Agent Wallet splits the difference. The user retains full control of the private keys and the Secret Recovery Phrase, while the agent operates from a dedicated transaction account. Before any transaction executes, the wallet checks it against user-defined spending limits, approved protocols, and operating policies. Any transaction that exceeds those thresholds or triggers a threat alert is held for human approval through 2FA.
This design is closer to a corporate expense policy than a personal wallet – the agent has execution authority within a fixed envelope, and everything outside that envelope requires a human signature. For crypto-native traders running automated strategies, that structure reduces the blast radius of a compromised agent without requiring manual approval of every single trade.
Agent Wallet bundles several protection mechanisms into the execution flow:
The coverage applies across nine chains: Ethereum, Linea, Arbitrum, Avalanche, Optimism, Base, Polygon, BNB Chain, and Sei. That scope matters because automated trading strategies often span multiple L2s and sidechains; a single-chain protection layer would leave uncovered positions elsewhere.
MetaMask splits the user experience into two modes. Guard Mode is the default: policy enforcement is active, flagged transactions route to the human, and approval workflows are required for any out-of-policy action. Beast Mode reduces the friction by removing some approval gates, aimed at users who prioritize speed and automation over strict oversight.
The practical takeaway: Guard Mode is the appropriate choice for agents managing meaningful balances or executing untested strategies. Beast Mode may suit low-value, high-frequency operations where the cost of a false-positive human review exceeds the potential loss from a single rogue transaction.
The wallet launched through a limited Early Access Program, with broad public availability expected later this summer. It supports leading agent frameworks including OpenClaw, OpenAI Codex, Claude Code, Nous Research Hermes Agent, and Cursor, which covers the majority of agent workflows currently being deployed onchain.
The near-term catalyst is the mass public rollout, which will determine whether Agent Wallet gains adoption beyond the early-access developer crowd. Two questions follow: whether competing wallet providers (including non-custodial options like Rabby and Frame) introduce similar agent-specific features, and whether the coverage cap or supported chain list expands before the public launch. Any changes to either would signal MetaMask's market assessment of where the real adoption risk lies – execution loss or chain fragmentation.
For now, Agent Wallet gives traders a structured answer to the question of how to let an AI handle onchain activity without handing over the keys. That alone moves the conversation from "should agents trade?" to "how do we bound their trading?"
Prepared with AlphaScala research tooling and grounded in primary market data: live prices, fundamentals, SEC filings, hedge-fund holdings, and insider activity. Each story is checked against AlphaScala publishing rules before release. Educational coverage, not personalized advice.