Skip to main content

How it works

Wallet for Agents has two sides. People pilot it from the Keyban dashboard: they set the limits, fund the wallets, approve what needs a person and follow the spending. The agent operates it from where it runs, through the Keyban MCP server or CLI: it pays, waits when a payment needs a person, and reports a refusal. Keyban sits between the two and checks every payment against the limits before it co-signs.

  1. CFO, admins, operators

    People pilot, in the dashboard

    • Set the limits

      Budgets, where the agent may pay, approval threshold

    • Fund the wallets

      Card, USDC or the multisig treasury

    • Approve and pause

      Held payments in the Inbox, emergency stop

    • Follow the spending

      Expenses by wallet, operator and mission

    agent

    The agent operates, through the MCP or CLI

    • Learns the wallet

      From the Keyban skill

    • Pays x402 services

      On its own, within the limits

    • Waits when needed

      A payment above the threshold goes to a person

    • Reports refusals

      With the reason, never works around them

  2. Keyban, in between

    • Checks every payment

      Against the limits, then co-signs what passes

    • Asks a person when needed

      Held payments land in the Inbox

On one side, people pilot from the dashboard: they set the limits, fund the wallets, approve and pause, and follow the spending. On the other side, the agent operates through the Keyban MCP server or CLI: it learns the wallet from its skill, pays x402 services, waits when a payment needs a person, and reports refusals. In between, Keyban checks every payment against the limits, co-signs what passes and sends what needs a person to the Inbox.

In short, people decide and watch from the dashboard, the agent pays from where it runs, and Keyban enforces the limits in between.

The people who pilot are the organization's Owners and Admins, often the CFO, who set the limits, and the operators of each wallet, who approve held payments and can pause it. They can also pilot from their own AI assistant instead of the dashboard, through the Keyban MCP server, with their own sign-in and rights: see Admin tools. They sign in on their own machine, never on one where an agent runs.

How the money flows​

You fund an agent wallet with a bank card, or by sending USDC you already hold. The agent spends that USDC on x402 services.

  1. Card or existing crypto

    Buy USDC with a card through the Stripe onramp, or send USDC you already own.

  2. Agent wallet

    One wallet per agent, with its own budget, approval threshold and emergency stop.

  3. x402 services

    The agent pays online services on its own, within the limits you set.

Money enters from a card, through the Stripe onramp, or from crypto you already hold; it lands in the agent's wallet; the agent spends it on x402 services.

In short, money enters by card or as USDC you already hold, lands in the wallet of one agent, and the agent spends it on x402 services.

Funding details to know before you start:

  • Card top-ups run through Stripe Link, on Base mainnet only, for between 5 and 999 euros per purchase.
  • x402 payments work on Base mainnet and XDC mainnet.
  • A new account can receive a Keyban welcome offer: a small USDC credit, granted when you create your first agent wallet on Base, that covers x402 payments for a limited time. It is not guaranteed, because it depends on the offer still being available when you create the wallet.

A fleet of agents can also be refilled from one shared pot that several people sign off: see Multisig treasury.

Two ways to use it​

You can give a wallet to the assistant you work with, or build it into agents that run on their own. Both are used in production, and one organization can do both.

  1. a person asks

    With your assistant

    • Claude, ChatGPT, Claude Code…

      The assistant you already use

    • Connected in one line

      From the dashboard's Connect the agent

    • Pays when you ask

      While you work with it

    no one in the conversation

    In the agents you build

    • Claude Platform, LangChain, CrewAI…

      Agents that run on their own

    • Integrated by your coding agent

      In your repository, with the Keyban CLI

    • A wallet per agent, team or customer

      Each with its own limits

  2. Keyban, for both

    • Checks every payment

      Against the limits you set

    • One dashboard

      Budgets, approvals, emergency stop, spending

With your assistant: the assistant you already use, such as Claude, ChatGPT or Claude Code, is connected in one line from the dashboard and pays when a person asks it. In the agents you build: agents that run on their own on Claude Platform, LangChain, CrewAI or another runtime, integrated by your coding agent with the Keyban CLI, each wallet standing for an agent, a team or a customer. For both, Keyban checks every payment against the limits and people pilot from the dashboard.

In short, an assistant that pays when you ask is a complete use on its own, and agents you build get their wallets through your coding agent; Keyban applies the same limits to both.

  • With your assistant. A person works with their assistant, which pays the services the task needs. You connect it from the dashboard: see Get started. Missions let it repeat a task at a known cost.
  • In the agents you build. Your agents run on a platform with no person in the conversation. Your coding agent checks that it fits, creates the wallets with the Keyban CLI and hands each one to its runtime. You decide what a wallet stands for: an agent, a team or a customer. See the Integration guide.

Where each thing lives​

WhatWhereWho
Limits, funding, approvals, spendingThe dashboard, at https://admin.keyban.ioPeople
PaymentsThe Keyban MCP server or CLI, where the agent runsThe agent
What the agent knows about the walletThe wallet skill, https://docs.keyban.io/.well-known/agent-skills/keyban-agent-wallet/SKILL.mdThe agent
InstallationThe setup skill, reached from the line the dashboard givesThe agent
Piloting from an assistantThe Keyban MCP server, after the person signs inPeople
Managing from your own codeThe REST API, with an organization API key: see CLI, MCP and APIYour backend