What is Pocket
Give each agent its own wallet and a hard cap it cannot raise.
Your agents need to buy things. Search results, market data, compute, other agents' output. Most of that is sold per call, and none of it will take an invoice from a language model.
So you have a choice. Hand your agent a card and hope, or give it an account that cannot be talked into anything.
Pocket is the second option. Your agent gets its own wallet, its own budget, and limits it cannot argue its way past. It buys what it needs. You get a receipt for every attempt, including the ones you refused.
What you can do with it
Create an agent
Give it a wallet in about ten seconds. It starts unable to spend anything.
Set its budget
How much a day, how much per purchase, and who it may pay.
Let it buy things
Any seller that prices itself openly, not a catalogue we curate.
See everything it did
Every purchase and every refusal, with the reason attached.
The idea in one line
Your agent
asks to buy something
checks your limits
Wallet
signs, only if allowed
Your agent
gets what it asked for
Your agent asks, Pocket checks your limits, and only then does a wallet sign
The order matters. Pocket decides before anything is signed, so a refused purchase has no signature attached to it. There is nothing sitting around that a clever prompt could talk an agent into using.
Three things worth knowing
Nothing is allowed until you allow it. A new agent cannot spend a cent. Every permission is one you granted, on purpose.
Refusals are kept, not dropped. When Pocket stops a purchase it records what was asked for, what stopped it, and how much room was left. Weeks later you can still answer "why did my agent stop?".
Your agent never holds a key. It asks Pocket to pay. It cannot sign for itself, so a leaked prompt cannot spend anything.
Works with what you already run
Hermes, Claude Code, Codex, or anything else that speaks MCP. One command connects it, and there is no key in that command.