Market-structure
MetaMask opened general access to Agent Wallet on August 6, a self-custodial wallet built to let Claude Code, Codex, Cursor and other coding agents execute on-chain trades through a CLI, constrained by spending limits the user sets in advance
Agent Wallet gives an AI coding agent its own wallet rather than the user's private keys, then enforces those limits through simulation, threat scanning and a mode called Guard Mode that pauses anything outside policy for two-factor human approval. An opt-in Beast Mode trims those interruptions for users willing to trade oversight for speed.
The significant design choice in Agent Wallet is not that an AI agent can trade, plenty of crypto products already do that, it is that the agent gets its own wallet with its own limits rather than borrowing the user's keys. That separation is what makes the rest of the security model coherent, an agent operating inside a scoped wallet with a daily spend cap and an allowlist of protocols cannot exceed what it was authorised to do even if the underlying model misjudges a trade, because the constraint sits at the wallet layer rather than inside the model's own judgement. Users connect through familiar coding-agent tools, Claude Code, OpenAI Codex, Cursor, OpenClaw, Hermes and OpenCode, over a command-line interface, which puts this squarely in the hands of the developer-trader segment already comfortable scripting strategies rather than mainstream retail.
Guard Mode is the default and does the actual constraining, daily spend limits, an allowlist of protocols the agent can touch, and a hard stop requiring two-factor approval through email or the mobile app for anything that falls outside those rules. Every transaction, in either mode, passes through three checks before it settles, a simulation showing the expected outcome, a threat scan run through Blockaid, and MEV protection against the kind of front-running that has quietly taxed on-chain traders for years. Beast Mode, opt-in only, reduces how often the agent has to stop and ask, which is the trade-off MetaMask is explicitly asking users to make with eyes open rather than defaulting them into.
Equities do not settle on MetaMask's rails, so this product has no direct line into share trading today. What it does have is the clearest working example yet of the exact pattern that Interactive Brokers, Robinhood and others are separately building toward through the MCP standard, an AI agent that can read a portfolio and propose trades, with a hard boundary somewhere between proposal and execution that a human or a rule engine controls. IBKR's version keeps order submission outside the AI session entirely, MetaMask's version lets the agent submit directly but inside pre-set financial guardrails, two different answers to the same underlying question of how much autonomy an agent should get once its owner has set the rules.
The part worth watching is Beast Mode, because it is the one place in this design where the human-in-the-loop checkpoint is optional by design rather than structural. On-chain settlement is final in a way equities trades are not, there is no broker-side compliance desk to catch an error before it clears, so a wallet-layer control that a user can choose to loosen is a meaningfully different risk than a control the platform enforces unconditionally. MetaMask's bet is that traders who want the speed will accept the tail risk that comes with fewer stops; the honest test of that bet will be whichever Beast Mode account has its first expensive mistake, and how MetaMask's simulation and threat-scanning layers hold up when that happens.
Read the original: MetaMask - MetaMask opened general access to Agent Wallet on August 6, a self-custodial wallet built to let Claude Code, Codex, Cursor and other coding agents execute on-chain trades through a CLI, constrained by spending limits the user sets in advance. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.