MetaMask’s Agent Wallet lets AI agents trade crypto on a user’s behalf, and its headline safety feature is up to $10,000 a month in loss protection. That protection is Transaction Shield, a $
MetaMask’s new Agent Wallet advertises up to $10,000 a month in loss protection for AI-driven trades. That protection isn’t new. It’s Transaction Shield, a $9.99-a-month subscription MetaMask launched for its wallet generally on December 3, 2025, eight months before Agent Wallet existed, now applied to the AI-agent pitch as if it were built for this specific product.
Your agent finally gets its own wallet.
MetaMask Agent Wallet is now live for everyone. pic.twitter.com/F81PZ7etxv
— MetaMask
(@MetaMask) August 6, 2026
Agent Wallet itself is new: a self-custodial wallet that lets AI agents, running on frameworks like Claude Code, Codex, and Cursor, execute swaps, perpetuals, and DeFi trades on a user’s behalf, inside spend limits the user sets. It first opened as a CLI-only early access program in June 2026; August 6 marked its wider public launch, and general availability (GA) is planned for later this summer. Agent Wallet is also the latest piece of a broader build-out, detailed below, that’s been steadily attaching paid features to more of the wallet.
Consensys CEO Joe Lubin, in MetaMask’s own June announcement, framed the product as necessary infrastructure for an “onchain economy” that “won’t be driven by humans alone.” The $10,000 figure is doing a lot of work in that pitch, and it deserves the same scrutiny as any headline number in a funding or product story: what’s disclosed next to it, and what isn’t.
Transaction Protection, MetaMask’s name for what Shield actually delivers, isn’t bundled into Agent Wallet. It’s a separate subscription: $9.99 a month, or $99 a year with a 14-day free trial, payable in mUSD (MetaMask’s own stablecoin), USDC, USDT, or by card, cancel anytime. It caps out at $10,000 in covered losses across a maximum of 100 eligible transactions per month: roughly three a day, thin coverage for an agent built to trade automatically and often.
And it predates Agent Wallet by the better part of a year. MetaMask introduced it in December as a standalone feature for the wallet at large, not as something built for the specific risks of letting a language model trade on someone’s behalf. Applying it to Agent Wallet is a repackaging, not new engineering, and that distinction matters for a product whose entire pitch is that its safety architecture was built for autonomous agents specifically.
MetaMask’s own documentation doesn’t agree with itself on this point.
SourceShield-eligible networks listedCountTransaction Shield subscription pageEthereum, Linea, Arbitrum, Avalanche, Optimism, Base, Polygon, BNB Smart Chain, Sei9Agent Wallet supported-chains referenceSame nine, plus Monad, HyperEVM, MegaETH, Robinhood Chain13Agent Wallet’s total supported mainnetsAll 22 networks Agent Wallet runs on22Either way, well under two-thirds of the chains an Agent Wallet user can actually trade on carry the protection MetaMask is marketing as the product’s headline safety feature. MetaMask hasn’t addressed the discrepancy, and there’s no way to confirm which number is current without a new statement or a documentation update.
These exclusions mirror standard crypto-insurance terms, but they cut directly against the specific risk Agent Wallet introduces: an autonomous agent making trading decisions without a human reviewing each one, where a bad trade, not a stolen key, is the likelier failure mode.
Users have 21 days to file after a qualifying loss. MetaMask says it processes claims in about 15 business days and pays out in mUSD, its own stablecoin, at the market rate at time of payment, not in the asset that was actually lost. MetaMask hasn’t published any data on how many Shield claims have been filed or paid since the subscription launched in December, an eight-month track record that would tell users more about the protection’s real-world reliability than the $10,000 ceiling does.
The chain-count mismatch isn’t the only place MetaMask’s own materials disagree. Its Agent Wallet marketing page claims the product works across “Hyperliquid, EVM chains, and Solana.” Its technical supported-chains reference, the document that lists every chain ID Agent Wallet actually runs on, includes no Solana network at all, mainnet or testnet. That doesn’t mean Solana support doesn’t exist. It means MetaMask hasn’t documented it anywhere checkable, and the marketing page’s Solana claim is unconfirmed.
The company’s own language about the underlying risk is more candid than the loss-protection headline suggests. Zhen Yu Yong, MetaMask’s Senior Director of Product, said in June:
“You cannot guarantee an LLM won’t be tricked. Prompt injection is an open research problem, not a bug you patch once.”
Prompt injection refers to attacks that smuggle hidden instructions into the data an agent processes, tricking it into acting against the user. Yong’s statement is a direct acknowledgment, from inside the company building the product, that the security model isn’t solving that underlying vulnerability. It’s building guardrails around a problem MetaMask itself says it can’t fix. Read against the loss-protection pitch, that puts Transaction Shield in a different light: not a guarantee that agent-driven trading is safe, but a financial backstop for a risk MetaMask has already said it can’t eliminate.
Agent Wallet is still in Early Access. General availability is planned for later this summer. That gives two concrete things to track. First, whether Shield’s eligible-chain list grows to match Agent Wallet’s own 22-chain footprint by GA. If it does, the current gap looks like a rollout lag. If it doesn’t, it supports reading Shield as a pre-existing revenue product attached to the AI story rather than infrastructure built for it. Second, whether MetaMask reconciles the conflicting chain counts, and whether it ever says how, or if, Shield’s terms were adapted for agent-driven trading risk rather than the general risk it was built to cover in December.
