Coinbase AgentKit Explained for Founders: What It Does and the Risk
AI Agents · 2026-11-04 · 3 min read
Coinbase AgentKit is a free, open-source toolkit that gives an AI agent a crypto wallet and a set of onchain actions, such as sending tokens. You write the agent. AgentKit supplies the wallet and the tools. The big catch is safety: its own README says it has no built-in approval step or spend cap.
Here is what it does and what you need, from the project's own README.
What is Coinbase AgentKit?
The README calls AgentKit Coinbase Developer Platform's toolkit for giving AI agents a crypto wallet and onchain interactions. It says you can use it with any AI framework and any wallet. It also says AgentKit is actively being built out.
Which wallets and frameworks does it work with?
The README's supported list shows three wallet options: Coinbase, Privy and Viem.
The README also links extensions for several agent frameworks: LangChain, the OpenAI Agents SDK, Strands Agents, the Vercel AI SDK and the Model Context Protocol (MCP), a standard for connecting tools to AI models.
For more on the wallet side, see our post on embedded wallets vs smart wallets.
What do you need to try it?
The README's quickstart lists these prerequisites:
- Node.js 22 or newer, or Python 3.10 or newer.
- A Coinbase Developer Platform secret API key.
- An OpenAI API key.
The quickstart's sample session asks the agent to fund its wallet with test tokens, and the Python example runs on the Base Sepolia test network. That means no real money is involved while you learn.
The code is free to use and change under the Apache-2.0 licence. We don't quote prices for the API keys. Check each provider's pricing page.
What is the big risk?
The README has a section called Managing Risk. It says AI models don't reliably tell instructions from data. So text that reaches the model can influence what the agent does, including transfers. It says this can't be fixed inside the toolkit.
It then says AgentKit has no human approval step, no spend cap and no list of allowed payees. Managing that is the developer's job.
In plain words: a funded agent can be steered. If someone tricks the model with bad text, the agent could move the money. So you add the limits yourself: a small balance, a cap per transfer, a list of allowed destinations and a person to approve big moves. The README points to guardrails middleware and an example that adds human approval for transfers.
Should a founder use it?
Use it if you have a developer and want control. Use a managed wallet if you want limits set for you. Our guide to stablecoin payments covers the payments side. More posts are on the blog index.
If you'd rather have a team build it, Devwiz builds AI agent products. To build one yourself, see how to build an AI agent with Claude.
What to do next
Write the three limits you'd set before real money goes in: the most the wallet holds, the most it sends at once, and who it may pay.
Frequently asked questions
Is Coinbase AgentKit free?
The code is free under Apache-2.0. You'll still need an API key from Coinbase and one for your AI model. Check their pricing pages.
Can I try it without risking money?
Yes. The quickstart uses a test network and test tokens. Write down how long it takes and which step needed a doc you couldn't find. That list is your real setup cost.
Does AgentKit work with MCP?
The README links an MCP extension. It also lists an MCP example in its repo.
Is it safe to give an agent real funds?
Not by default. The README says there is no approval step, spend cap or allowlist built in. Add your own.