Detect transaction intent. Prove every action.
After an agent acts, no verifiable record exists. ZeroProof issues a receipt for every tool call, signed by both the agent and the tool server, anchored on-chain.
An agent can call the wrong tool and still sound certain. ZeroProof's first model is a small transformer, post-trained for transaction intent. It flags tool calls that drift from user intent before they run.
We fine-tune small, task-specialized models to read what a person actually means before an agent acts on it. Ecommerce Intent 1B is our first, focused on payment and commerce conversations. At around a billion parameters it is small on purpose, cheap enough to run on every action instead of only the flagged ones, where a frontier model would be too slow and too costly. On our balanced eval it routes intent at a level competitive with much larger models.
ZeroProof is a model lab building the trust layer for AI agents. It post-trains models to read user intent, then proves every action with a verifiable receipt.
An MCP-connected agent sends a tool request. ZeroProof intercepts it before execution.
Before it runs, the intent model checks the call against stated user intent.
If intent is confirmed, a bidirectional cryptographic receipt is issued, signed by both the agent and the tool server.
The verified action is anchored on-chain. Each receipt contributes to a tamper-proof, auditable agent reputation.
ZeroProof is in early access. Get notified when the intent model and Python SDK ship.