Inspiration
We were excited to hear about how AI Agents were being used for general-purpose everyday tasks such as booking appointments and planning vacations but noticed that people were (rightfully) skeptical in trusting agents with their money.
What it does
Our MCP sits between an agent and a payment provider (such as Stripe) and rigorously vets the AI's spending.
Features
- Two-level vetting.
- First the whole plan is vetted: the agent breaks the goal into a graph (item -> subgoal -> goal) with a price on every node. Then each individual purchase is vetted against that accepted plan.
- AI judges plus deterministic rules.
- Provable guarantees for anything that is arithmetic or a lookup: a node's price is the sum of its children, the total and cumulative spend stay within budget, an item can't cost more than its estimate, etc.
- Strong empirical guarantees from an LLM judge (Jev) for anything that needs judgment: whether this is the right item, whether the URL is a legitimate merchant or a phishing site, and whether each step serves its parent.
How we built it
- Fullstack Framework: Jac
- Hosting: Render.com
- LLM: Gemini
- Judge: Jev (typesafe.ai)
Challenges we ran into
- Free-tier small agents (14-30B parameters) hallucinating payment URLs instead of searching the web.
- Websites blocking visits from AI Agents
Accomplishments that we're proud of
- Fully shipped MCP that you can use with ClaudeCode or Codex or any other AI Agent today
- Official Stripe Mock API Integration
- Latest state-of-the-art models (Jev) as a Judge (Extremely fast, cheap, and accurate)
- 100% Jac and CSS coded
What we learned
- Writing MCPs
- Working with Agentic AI systems
- Working with Jaclang
What's next for CounterSign MCP
- More testing, especially in rogue agent cases
- A real-world test with actual money involved (to prove that CounterSign let's us trust agents with our money)
- Solid metrics or evidence for our claims
Built With
- gemini
- jac
- render
Log in or sign up for Devpost to join the conversation.