Security & privacy
What the server can read, what its token can do, and what the model is not allowed to do.
What it can read
Every tool reads public data: hospital price files, insurers' published rates, CMS and Medicaid volumes, and Medicare payment tables. None of them reads your team's uploaded files, contracts, claims or studies in CognitoLabs, so connecting the server to a client engagement exposes nothing your client gave you.
What the token can do
- The access token it issues works on the MCP server only. Our regular API refuses it, and the MCP server refuses tokens from the regular API.
- The token never carries admin rights.
- It lasts 2 hours. See Signing in.
What you approve
Before a client is connected, the consent page shows the client's name and the host it will send you back to, and the account it will act as. Clients can only return to an https address or to your own machine. You can deny.
No SQL from the model
Each tool takes typed filters: a state, codes, hospital ids, a short search text. The model never writes SQL or any query language that runs against our data, so there is nothing for a prompt to inject into. Bad input comes back as a plain error.
What we record
For each tool call we keep who made it, which AI client it came from, the inputs the model sent, and a shortened copy of what came back. We use it for support and to improve the quality of the answers. We also count calls per person to apply the daily limit.
If you tell the assistant an answer was good or bad and agree to send feedback, we also receive your rating, your comment, your question and a short summary of the answer. The assistant should never include patient names or other personal health information, and you can say no.
We never see the rest of your chat with the assistant.
Limits
Registration and sign-in requests are limited to 20 per minute per IP address.