Skip to main content
Agents that move money need two things a human doesn’t: a hard budget, and machine-readable docs. Both are built in.

​
A key the agent can’t overspend

Give each agent its own API key:
  • Scopes limit what it can do — payouts:write and recipients:read for a supplier-paying agent; nothing else.
  • Spend caps limit how much — per transaction, per day, per month, in USD across all currencies. A payout that would exceed a cap is rejected before any money moves, whatever the agent’s reasoning was.
  • IP allowlists limit from where.
  • externalReference makes every money-moving call idempotent, so an agent that retries after a timeout can’t double-pay.
The blast radius is then a number you chose in the dashboard, not a property of the agent’s behaviour. Revoke the key to stop it instantly.

​
Machine-readable everything

  • Every page of these docs is available as Markdown: append .md to any URL, or fetch /llms.txt (index) and /llms-full.txt (everything).
  • The docs are also an MCP server — connect an agent (Claude Code, Cursor, any MCP client) and it can search and read this reference directly. See the Connect option in the page menu.
  • The API itself is friendly to agents: stable error codes to branch on, details with structured context, ids that announce their type, and the capabilities catalog to discover what’s possible at runtime instead of guessing.

​
A sane loop

  1. GET /v1/capabilities/{country} — what can I do here?
  2. POST /v1/validate — is this destination real, and whose is it?
  3. POST /v1/payouts with an externalReference derived from the task id.
  4. Wait for payout.completed / payout.failed — don’t poll in a tight loop; rate limits apply to agents too.
  5. Log the x-request-id of every call. When something looks wrong, that’s what support needs.