Test a model on data you never see.
Test an AI model on data you can’t access, and share data without giving it away. Only the score comes out.
See a real paid run on NEAR testnet- NEAR Protocol
- OutLayer
- TDX
- x402
How it works
You’re building an AI model. You don’t have the data.
To prove it works, you need real records. Patients, transactions, customers.
Someone has it. They can’t share it.
A hospital’s records are protected by law. Sending them to you isn’t an option.
Now you can test on it anyway.
Your model runs on their data inside a hardware enclave (a TEE): a sealed part of the chip that no one can look into. You get the score. Neither side sees the other’s file.
You’re building an AI model. You don’t have the data.
To prove it works, you need real records. Patients, transactions, customers.
Your model
You
Patient data
A hospital
No data to test on
Where it’s useful.

A hospital
Patient records
A diagnostics startup
Diabetes-risk model
“Does the model spot diabetes risk in our patients?”
Agents can pay for it too.
Add the SecureLambda MCP server to your coding agent. It finds a dataset, pays per run in NEAR over x402, and reads back the score.
claude mcp add securelambda -- npx -y @securelambda/mcp-serverThe agent pays from a wallet you fund, so it can’t spend more than you sent.
Setup, spend caps and plain HTTP in the docsTest my churn model on telecom data. Spend at most 0.05 NEAR.
- securelambda · list_listingsTelecom churn · 0.02 NEAR per run
- securelambda · quote_evaluation402 · 0.02 NEAR to the data owner
- securelambda · evaluatePaid 0.02 NEAR · scored in the enclave
Accuracy 0.79 on the operator’s data. Their data never left the enclave. Receipt is on NEAR.
Questions
Your files are encrypted in your own browser before upload and are only opened inside the hardware enclave. The enclave’s key is set up once by us and our copy deleted, so you’re trusting that step. The next version has the enclave create its own key, so nobody ever holds it.