AI Agent Wallets Compared: Who Holds the Keys and Who Sets Limits
AI Agents · 2026-11-21 · 5 min read
An AI agent wallet is a crypto wallet that software can use, so an AI agent can pay for things without a person clicking approve every time. Two questions matter most: who holds the keys, and who sets the limits. The managed products from Coinbase, Privy and Crossmint answer both. A raw toolkit may answer neither.
This post compares what each provider says in its own docs. No rankings. No buy advice.
What is an AI agent wallet?
It's a wallet with an agent as the user. The agent can sign payments, pay for web services or move stablecoins within rules you set.
It matters because agents now buy things: data, compute, tool calls. Card checkouts are built for people. A wallet lets software pay in seconds.
For the business view of taking these payments, see our guide to stablecoin payments for small business.
How do the main options compare?
Here's the summary from each provider's docs, read on 22 November 2026.
| Option | Who holds the keys | Spend limits | Notes |
|---|
| Coinbase Wallet MCP | A dedicated agent wallet with an expiring signing grant. Not your main wallet | Saved spending controls | Also has an approval mode, where you sign each action |
| Privy agentic wallets | Your backend (developer-owned), or the user, with the agent as an extra signer | Policies, set per wallet | Two control models in its docs |
| Crossmint agent wallets | The user, or your backend. Crossmint says neither it nor you takes custody of funds | Spend cap, a list of who the agent may pay and a time window | Limits are enforced onchain |
| Coinbase AgentKit (open source) | You choose a wallet provider | None built in | The README says it doesn't enforce caps |
These are the vendors' own descriptions, not audits.
What does Coinbase say?
Coinbase's Wallet MCP docs describe two ways an agent can act.
- Approval mode: your agent prepares an action for your main wallet. You review and sign through an approval link.
- Agent-wallet mode: a dedicated wallet and an expiring signing grant let your agent perform supported actions within saved spending controls.
The docs say agent-wallet mode does not give unrestricted control of your main wallet. They also say an agent wallet has its own address, balance, activity and spending policy.
What does Privy say?
Privy's agentic wallets recipe describes two setups.
- Agent-controlled, developer-owned. Your application backend controls the wallet through authorization keys.
- User-owned. The user keeps ownership and the agent acts as an extra signer with scoped policies. The user can revoke the agent's access at any time.
What does Crossmint say?
Crossmint's agent wallet docs describe a wallet where the owner holds the keys and sets the limits the agent can spend within. For a consumer agent, the user owns the wallet and adds the agent as a signer, with a spend cap, a list of who the agent may pay and a time window. For an autonomous agent with no end user, your backend owns the wallet.
The docs say permissions are enforced onchain, so a transaction outside the limits is rejected at the wallet level, and the owner can revoke them at any time.
What about raw toolkits?
Coinbase's open-source AgentKit gives an agent a wallet and onchain actions. Its README warns that it doesn't gate transfers behind human approval, enforce spend caps or allowlist destinations. So if you use it alone, you must add those limits yourself.
How do agents pay for web services?
Many use x402, a free open standard that lets software pay a website directly. The x402 docs say the standard itself has 0 fees built in. A wallet or service you use with it may still charge, so read its pricing.
How should a founder choose?
Ask in this order.
- Who holds the key? If your backend does, you carry the security job.
- Where do the limits live? Prefer limits enforced by the wallet, not only in the prompt.
- Can a human step in? Set a level above which a person approves.
- What does it cost at your volume? Read the provider's pricing page.
For the risks of agents with wallets, read our post on DeFAI. For the human side of wallets, see embedded wallets vs smart wallets. Want to build the agent yourself? Start with this guide to building an AI agent.
What to do next
Write down the three limits for your first agent: the most it can hold, the most it can send at once, and who it may pay. Pick the provider that lets you enforce all three.
Frequently asked questions
Is an AI agent wallet safe?
It depends on the limits. Providers describe spend caps, policies and expiring grants. Those reduce risk, not remove it. Start small.
Can the agent steal my funds?
An agent can only spend what the wallet lets it. Set caps, an allowlist and a human check above a limit.
Do I need a blockchain for an agent to pay?
Not always. Card networks have agent products too. Wallets and stablecoins are one route, and a good fit for small, fast payments.
How do I test a wallet's limits?
Fund one wallet with a small amount and set a low cap. Ask the agent to make one payment inside the cap and one outside it. The second must fail. If it doesn't, the limit isn't enforced where you think.
What does it cost to try?
We don't quote prices here. x402 as a standard has no fee. Check each provider's own pricing page.