Mandate AI

light mode
Command tab
forest mode
Mandate tab
Approval tab
Approval tab
fair share tab
fair share tab
Spend Firewall Simulator (tools tab)
Agent Playground (tools tab)
Attack Lab (tools tab)
red team scorecard (tools tab)
tamper evident ledger (tools tab)
Agent Marketplace (tools tab)

Inspiration

AI agents are starting to spend real money: buying cloud compute, paying vendors, settling invoices. But handing an agent a raw card or bank account means hallucinated purchases, runaway API bills and malicious vendors, with no scoped authority. We wanted agents to be able to pay without being able to ruin you.

What it does

Mandate is a spend firewall and finance console that sits between AI agents and PayPal. Users issue scoped mandates: budget limits, per-transaction caps, merchant allowlists and expiry dates. Every payment request goes through a deterministic policy engine with three outcomes:

Auto-approve when it fits the mandate, and PayPal settles it
Escalate to a human approvals queue above the threshold
Hard block when it violates the merchant list or caps

Beyond the firewall, Mandate includes:

FairShare: a conversational group expense splitter. Upload receipts, type "Alice didn't drink, Bob left early, Charlie covered the deposit", and it assigns line items, calculates exact shares in integer cents and generates PayPal payment links.
Dispute Resolver: drafts merchant rebuttals from order and tracking evidence and estimates recovery odds.
Procurement Agent: matches invoices to purchase orders, flags anomalies and requires sign-off before payout.
Freelancer CFO: invoicing, late-payer follow-ups and multi-currency summaries.
A real-time decision preview, an audit log of every decision, and Light / Dark / Forest themes.

How we built it

Backend: Node.js + Express 5 using the official @paypal/paypal-server-sdk (Orders: create and capture) in the PayPal Sandbox.
Frontend: React + Vite with a custom responsive design system.
AI: Google Gemini for receipt understanding, natural-language split notes and drafting.
Safety by design: the policy engine is deterministic, so AI can assist but never override the rules. Secrets stay in environment variables.

Challenges we ran into

Making natural-language notes map reliably to line items while keeping the math exact to the cent.
Designing an approval flow that is strict without making demos slow or confusing.
Keeping a large feature set cohesive, responsive and consistent across three themes.

Accomplishments that we're proud of

Real sandbox settlements end to end, a firewall that blocks bad payments with clear reasons, and a polished UI that works on any screen size.

What we learned

Agent commerce needs governance before autonomy. Clear limits and audit trails matter more than clever prompts.

What's next for Mandate

An MCP server so any agent can request payments through Mandate, a prompt-injection shield, a red-team scoreboard, mobile approvals via Telegram, and a tamper-evident audit ledger.