MetaMask launched Agent Wallet on August 6, 2026 — a self-custodial wallet built to let AI agents monitor markets and execute on-chain trades on a user's behalf, inside limits the user sets upfront. It follows a smaller June early-access run with roughly 200 users, per [Decrypt](https://decrypt.co/375044/metamask-self-custodial-ai-wallet-autonomous-crypto-trading). The pitch is the one every agent-wallet launch makes right now: let the bot act, but only inside a cage you build first.
That framing is the interesting part. Not "AI trades crypto" — every quant desk has done that for a decade — but a consumer self-custodial wallet deciding that agent permissioning, not agent intelligence, is the product surface worth shipping.
What Agent Wallet actually does
Before an agent gets to touch funds, a user configures:
- ▸Spending limits and an allowlist of approved protocols
- ▸One of two operating modes — Guard Mode and Beast Mode — trading off caution for autonomy
- ▸Transaction simulation, threat scanning, and MEV protection on each trade
- ▸Gas abstraction, so an agent can swap or transfer without holding the chain's native gas token
MetaMask's own framing: "The agent can act, but it acts inside the user's boundaries." It supports Hyperliquid and EVM-compatible chains, and connects to agent runtimes builders are already using — Claude Code, Codex, Cursor, OpenClaw, Hermes, and OpenCode. That last part matters more than the trading angle: this isn't a walled-garden trading bot, it's wallet infrastructure meant to be wired into agents that already exist.
MetaMask also advertises Transaction Protection coverage up to $10,000 a month for transactions that pass its security checks.
Claimed vs shown
Here's where the Trust Rule earns its keep. "Self-custodial" is a real, checkable architecture claim — the user holds the keys, not MetaMask. Spend limits, protocol allowlists, and simulation-before-execution are concrete permissioning controls, the same category of thing SatoHub tracks as wallet controls across the agent-wallet stack.
What's *not* independently shown yet: the $10,000 protection coverage is MetaMask's own claim, with no third-party audit or claims-payout data cited in the announcement. "Guard Mode" and "Beast Mode" are marketing labels for automation levels — useful shorthand, not a security standard. And MEV protection, threat scanning, and simulation are described, not benchmarked against a public track record. None of that means the product doesn't work as described. It means a builder evaluating it should read "MetaMask says" where the announcement implies certainty, and ask for the audit or the incident history before wiring real funds through it.
That distinction — self-reported feature vs. independently verified behavior — is exactly the gap a Sato Score exists to surface for any wallet or tool in the builder library: not a verdict on whether Agent Wallet is safe, just a transparent accounting of what's evidenced and what isn't yet.
Why this is a category signal, not just a product launch
MetaMask isn't alone. Coinbase and MoonPay have both moved on AI agent wallet infrastructure in 2026, per the same report. When three of the biggest names in consumer crypto wallets converge on "spend limits + allowlists + simulation" as the baseline for letting an agent hold keys, that's the industry quietly agreeing on a permissioning shape — not autonomy without limits, autonomy *inside declared limits*. That's the same principle SatoHub's builder maps onto every wallet skill in the stack: an agent that can move funds needs a spend cap and an allowlist before it needs a bigger model.
What to watch
Watch for: an independent audit of Agent Wallet's simulation/threat-scanning pipeline, real numbers on the Transaction Protection payout (not just the ceiling), and whether Guard Mode/Beast Mode settings get standardized language other wallets can compare against. If MetaMask publishes any of that, it moves from claimed to shown — and that's the marker worth tracking, not the trading headline.
If you're building an agent that needs to hold or move funds, the permissioning shape matters more than which wallet ships it first — check what spend-limit, allowlist, and simulation controls are actually available before you wire one in.
Sources
- ▸[MetaMask Launches Self-Custodial AI Wallet for Autonomous Crypto Trading — Decrypt](https://decrypt.co/375044/metamask-self-custodial-ai-wallet-autonomous-crypto-trading)