Agentra

• 

  Welcome page

• 

  Welcome page

• 

  Homepage / Dashboard

• 

  Chat page

Inspiration

What it does

How we built it

Inspiration

AI agents can already search, analyze, and complete tasks, but letting them spend money is a different level of trust. We built Agentra after asking a simple question: How can an AI agent purchase the services it needs without receiving an unlimited company card?

What we built

Agentra is a control layer for AI-agent purchases. An agent can discover and compare paid APIs, while Agentra checks spending limits, service categories, and approval requirements.

If human approval is needed, the process pauses before payment or provider execution. After approval, Agentra demonstrates the x402 payment flow, calls the service, and creates a clear receipt showing what was selected, why it was allowed, who approved it, and what happened.

We also created a Policy Simulator so teams can test new rules against historical transactions without moving funds or changing existing records.

How we built it

We used Next.js, React, TypeScript, PostgreSQL, Prisma, Redis, Auth.js, Docker, and OpenTelemetry. The platform supports team workspaces, role-based permissions, API providers, transaction history, policy controls, approvals, and simulated x402 payments.

A simplified policy decision looks like:

[ \text{Allow} = (\text{Cost} \leq \text{Transaction Limit}) \land (\text{Daily Spend} + \text{Cost} \leq \text{Daily Limit}) ]

A purchase above the approval threshold is paused until a human authorizes it.

Challenges

The hardest challenge was making the agent useful without allowing it to bypass human control. We had to ensure that the final provider result appeared only after approval and that retries could not create duplicate payments.

We also worked through multi-workspace data isolation, historical policy calculations, responsive transaction tables, accessibility, and reliable local infrastructure with PostgreSQL and Redis.

What we learned

We learned that agent payments are not only a payment problem. They are also a governance, security, and communication problem. Teams need to understand why an agent made a purchase—not just whether the payment succeeded.

That insight shaped Agentra around three ideas: discover carefully, approve responsibly, and document everything.

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Agentra

Built With

• auth.js
• docker
• next.js
• node.js
• oauth
• ogl
• opentelemetry
• playwright
• postgresql
• prisma
• react
• recharts
• redis
• rest-api
• tailwind-css
• typescript
• vitest
• webgl
• x402
• zod