What one AI request goes through before it leaves your company
One prompt, six steps, and a record you can walk someone through afterward. Governance is the point. A smaller token bill is the bonus.
A walk through the Kairos gateway
A single AI request looks like a question in a box, typed by someone who is still in the building. The moment it needs a model, it is on its way out of the company. What it carries, and what you can show later, depends on the path it takes before a provider sees it.
Follow one request, the kind the Kairos homepage already walks through. A teammate asks for help with a note. In the prompt are a patient name, a Social Security number, and an API key pasted in because that was the fastest way to pass it along. The site names the finding directly: three personal data fields and one API key in the prompt.
Kairos meets that request at an OpenAI-compatible base URL, https://api.kairos-ctx.ai/v1, with a kai_ key. Editors, scripts, applications, and agents can share that endpoint. From there the call follows the homepage order: Inspect, Redact, Compress, Check policy, Route, Record. Governance is why the path exists. A shorter prompt and a careful route can lower the bill, and that is a bonus. The part you keep is the account of this request. Evidence, not assurances.
1. Inspect
On this call: three personal data fields and one API key, found in the prompt.
Inspect runs first, while the request is still yours. The gateway reads the prompt before any provider does. On this call it finds three personal data fields and one API key. You can say what was in this prompt, on this send, and every later step has something concrete to act on.
The person who pressed send stays in the tool they were already using. Inspection rides along with the call. A cleaner prompt on another afternoon would still be read. This one carries a patient name, an SSN, and an API key, so the path has all three in view before anything leaves. Look, name what you found, and pass a known request forward.
2. Redact
On this call: the patient name, the SSN, and the API key are replaced before anything leaves.
Redact changes the text. The patient name, the SSN, and the API key are replaced before anything leaves. The model can still be asked for help. The provider receives the request with those values replaced. Personal data fields and the key are handled in the same pass, because a secret pasted into a prompt travels the same way a name does: inside the body of the call.
This is the heart of the gateway. The patient name, the SSN, and the key are replaced before anything leaves. Sensitive values go first. Compression and routing come after, on a prompt that has already been through that replacement. Governance is the product. The savings happen beside it.
3. Compress
On this call: context trimmed from 12,480 tokens to 9,110, about 27%.
After those values are replaced, the prompt can still be heavier than the question needs. Earlier turns, repeated instructions, and context from a previous message all come along. Compress trims that extra context. On this call it goes from 12,480 tokens to 9,110. About 27% of the prompt is removed on this call. Past this one example, the site says Kairos compresses at least 25% of input tokens before they are sent on.
The 27% belongs to this call. At least 25% of input tokens is the wider claim, and both figures are about the size of the prompt. Redaction has already handled the patient name, the SSN, and the API key by the time the token count drops. You still get the help you asked for, and you send fewer input tokens to get it. The smaller prompt is a bonus.
4. Policy
On this call: 22 frameworks evaluated, and the request cleared.
Check policy comes next, between compression and routing, which is where the homepage places it. Twenty-two frameworks are evaluated on this request. On the walkthrough the site shows, those frameworks clear, and the call continues. The same prompt that was inspected and redacted is the prompt the frameworks are evaluated against.
When someone asks whether policy ran, you can point at this send. The number to cite is the one the site publishes: 22 frameworks, evaluated on the request and enforced on the way out. The result stays with that call, which is a sturdy thing to carry into a review.
5. Route
On this call: matched to the cheapest capable model, after the checks have run.
Route is the step that talks to a model. The request goes out through the same base URL, https://api.kairos-ctx.ai/v1, under the kai_ key the client already sent. The homepage describes this moment as matching the call to the cheapest capable model. A less expensive match is a welcome outcome, in the same family as compression: a savings bonus on a request that has already been governed.
What leaves is the prepared prompt. The patient name, the SSN, and the API key have been replaced. The context is the trimmed one, 9,110 tokens where 12,480 stood a moment ago. Twenty-two frameworks have been evaluated. The application still receives a normal completion. Routing chooses a model once inspect, redact, compress, and policy have done their work.
6. Audit trail
On this call: the decision is recorded. The site calls this step Record.
Record closes the loop. The decision is written to the audit trail, so you can show the path when someone asks what left the company. Inspect found three personal data fields and one API key. Redact replaced the patient name, the SSN, and the key before anything left. Compress took this call from 12,480 tokens to 9,110, about 27%, beside the wider claim of at least 25% of input tokens. Policy evaluated 22 frameworks, and they cleared. Route sent the governed request on. Six steps, one request, kept together.
An assurance is a sentence about how a company means to behave. A trail is an account of this call. Evidence, not assurances. When the question arrives, you can walk the same steps in the same order: the finding, the replacement, the smaller prompt, the policy check, the route, and the record that holds them.
Send the next one through the same path
The free plan is free forever. Set the base URL to https://api.kairos-ctx.ai/v1, use a kai_ key, and send one request of your own through Inspect, Redact, Compress, Check policy, Route, and Record. Open pricing is on the site if you want to see the plan before you create an account.
Prefer to watch the bonus on a key you already pay a provider for? The homepage includes Use my key. Bring that key and send a prompt. Governance is why the gateway sits in front of the request. Token savings are what you notice once the path is in place.