Mandate: The Bank for AI Agents

AI agents can work but can't borrow. Mandate is the bank for AI agents: underwritten credit against invoices, funded by other agents, settled on PayPal, with a human-signed mandate on every loan.

Comment

INSPIRATION:

AI agents already do real work, but they can't borrow. An agent that finishes a \$400 job but needs \$10 now to deliver it has one option: a human's credit card. So we built the lender agents don't have.

WHAT IT DOES:

Mandate is a sandbox bank for AI agents, and every \$ moves on PayPal.

  • Borrow: \$10-\$500 against a client invoice, paid out via PayPal Payouts.
  • Lend: Agents fund loans via PayPal checkout and earn 80% of the fees.
  • Repay: The client pays a PayPal link, webhooks fire, and the loan settles itself.
  • Any agent: Claude Code, Codex, Cursor, Gemini, Copilot. Every op is a tool over MCP, REST and OpenAPI, and each reply says what to do next.
  • Human control: The owner verifies with PayPal and signs a mandate capping the loan size, total owed, fee and term.
  • Rules decide, AI explains: A deterministic score gives grades A-E, priced as

$$\text{fee} = P\,(r_{\text{funds}} + \text{EL} + c_{\text{ops}} + c_{\text{cap}})$$

e.g. \$10 for 7 days at grade D costs \$0.58, of which lenders get \$0.46 and the bank \$0.12.

  • Collections: A late fee, then a penalty, a freeze and a charge-off. Losses are split pro-rata, $\text{loss}_i = L \cdot n_i / \sum_j n_j$, and the human gets a PayPal guarantee link.
  • Risk desk: Elastic ES|QL fraud detectors, Jina semantic search and AI case files.

HOW WE BUILT IT:

  • Stack: Next.js 16, TypeScript, Postgres, Drizzle and a double-entry ledger, on Render.
  • PayPal: Orders v2, Payouts, webhooks and Verify-with-PayPal. Request IDs give exactly-once effects, and there are never PayPal calls inside DB transactions.
  • Sponsor tools:
    • APIMatic: Its PayPal context plugin led to 5 real fixes, and it generated our SDKs.
    • Elastic + Jina: The risk desk.
    • Bryntum: The repayment timeline.
    • Postman: A public collection.
    • Kernel: Cloud browsers for agent tests.
    • AG Grid: The console's tables.
    • Render: Hosting, with auto-deploys.
  • Tests: About 1.1K unit and integration tests, E2E against the real PayPal sandbox, and real Claude Code agents running the full loop.

CHALLENGES WE RAN INTO:

  • Exactly-once money: Webhooks, crons and agents all race each other. Reviews caught a double-redemption race and payouts mis-booked as failed, and we fixed and tested each one.
  • Keeping the LLM out of decisions: A test feeds it prompt injections and checks that the decision stays the same.
  • Sandbox quirks: Payouts to unconfirmed emails went UNCLAIMED, so we pay to PayPal account IDs instead.

ACCOMPLISHMENTS THAT WE'RE PROUD OF:

Real Claude agents given only a URL opened accounts, got claimed by their human, borrowed, and repaid on the PayPal sandbox. The ledger balanced to \$0 every time.

WHAT WE LEARNED:

Credit for agents is classic lending discipline plus one new rule: every response must tell the agent what to do next. AI belongs in explaining decisions, not making them.

WHAT'S NEXT FOR MANDATE:

Real-money pilots with a licensed partner, agent-to-agent invoice factoring, and mobile approvals for mandate changes. Huzaifa previously built AI for micro-lending at easypaisa digital bank, one of Pakistan's leading digital banks. Mandate brings that playbook to AI agents, launching first with capped loans in a regulatory sandbox.

Built With

Share this project:

Updates

Submission history