What Is a Non-Custodial Wallet?
A non-custodial wallet is one where the keys that authorise transfers are controlled by you, not by a company holding funds on your behalf. No provider can move your assets without your signature. The trade is responsibility: if the keys are lost or stolen, there is usually nobody who can restore or reverse anything.
Custody is about who can sign
Onchain assets move when a valid signature authorises a transfer. Whoever controls the signing key controls the funds. That is the whole definition.
Custodial. A company holds the keys. You have an account balance in its system, and you ask it to send funds. Exchanges typically work this way. The company can recover your account if you forget a password, and it can also freeze, delay, or lose funds, and it is exposed to its own insolvency and security failures.
Non-custodial. You hold the keys, or a device or account you control does. The wallet software helps you build and sign transactions but cannot send anything without your authorisation. If the software company disappears, your assets are still on the chain and still yours, provided you still have the keys.
The app you use does not settle the question. The same interface can be custodial for one product and non-custodial for another. Ask where the signing key lives and who else can use it.
The middle ground
Some designs split signing authority, for example a key split across several parties with a threshold needed to sign, or a smart contract account that requires more than one approval. Whether these count as non-custodial depends on whether any single provider can move funds without you. Read the design, not the label.
How keys are held
Seed phrase wallets. A sequence of words encodes the root secret from which keys are derived. Whoever has the words has the funds. Simple and portable, and the main failure is the phrase being lost, photographed, or phished.
Hardware wallets. A dedicated device holds keys and signs internally, so the secret never touches a general purpose computer. Strong protection against malware on the host; still depends on the backup of the root secret.
Operating system keychains and secure enclaves. Keys stored in the platform's protected credential store, often backed by hardware, and released only to authorised processes. Well suited to software that signs on a machine, including agents.
Smart contract accounts. The account is a contract whose rules decide which signatures are valid. That allows spending limits, multiple signers, recovery by designated guardians, and scoped keys that can only do certain things. The account can be fully non-custodial while still being much harder to drain than a single key.
Each option moves risk around rather than eliminating it. The question is which failures you are better prepared to handle.
Non-custodial for agents
An autonomous agent that pays for things has to sign without asking a person each time. That makes the custody question sharper, because the agent's environment is exactly where a key is most exposed: it runs code, calls tools, reads untrusted input, and may be manipulated.
Do not give an agent an unrestricted key. A raw private key in an environment variable is non-custodial in the technical sense and a single point of total loss in practice.
Bound what the key can do. A scoped session key on a smart account can be limited to a spending cap, specific assets, specific recipients or contracts, and an expiry. If it leaks or the agent misbehaves, the loss is capped by design.
Keep the key in a protected store. An operating system keychain releases the key only to the authorised process, rather than leaving it in a file or variable any code can read.
Keep the owner key elsewhere. The key that can create, top up, or revoke session keys should not be available to the agent at all.
This is the model CryptoCadet uses: a non-custodial USDC rail on Base in which the agent spends from a float bounded by session keys, with custody held through the OS keychain. It is not a consumer wallet and does not hold funds on anyone's behalf.
What non-custodial does not protect against
It does not protect against signing something you should not have signed, against malware on the machine that holds the key, or against sending funds to the wrong address. Settlement is final. Non-custodial removes one category of risk, the provider, and makes the rest your responsibility to design for.
Frequently asked questions
- What is a non-custodial wallet in crypto?
- A wallet where you control the keys that authorise transfers, rather than a company holding funds for you. The wallet software builds and signs transactions but cannot move assets without your authorisation. If the keys are lost or stolen, there is usually nobody who can restore or reverse anything.
- What is the difference between custodial and non-custodial wallets?
- In a custodial wallet a company holds the keys and you hold a balance in its system, so it can recover access but can also freeze or lose funds. In a non-custodial wallet you hold the keys, removing provider risk and adding key management responsibility.
- Can a non-custodial wallet have spending limits?
- Yes. A smart contract account defines which signatures are valid, so it can enforce spending caps, multiple approvals, recovery guardians, and scoped session keys limited by amount, asset, recipient, and expiry, while no provider can move funds without the owner.
- Is it safe to give an AI agent a non-custodial wallet?
- Only with bounds. An unrestricted private key is a single point of total loss. A session key limited by spending cap, assets, recipients, and expiry, stored in an operating system keychain with the owner key kept elsewhere, caps the loss if something goes wrong.