Where your data goes, and what is written down.
Every statement on this page describes what the system does today, checked against the code that does it. Where something is not true yet, it says so rather than being written in the future tense.
Core providers used by the hosted service. Model providers and services connected by a customer also depend on that workspace’s configuration.
ECS Fargate runs the application; RDS Postgres holds tenants, policies, traces and the usage record and is not publicly reachable; Cognito authenticates console sessions; SSM SecureStrings hold secrets; CloudWatch holds application logs. Amazon Bedrock also provides managed model inference for configured models.
A model-routing service used alongside Amazon Bedrock. Inputs, selected images, conversation history and retrieved library content may reach a model through it. Every OpenRouter request asks OpenRouter for providers that do not train on prompts and for zero-data-retention endpoints only, on every workspace, with no setting to turn it off. A model no provider serves that way is refused when it is chosen, not on the turn it would have failed. On a bring-your-own-key workspace this is your account and your contract, not ours.
Holds the subscription outright — checkout, invoices, payment methods, cancellation. Card details never reach us. No mirrored copy of subscription state is kept here.
An agent turn can involve multiple model calls. Guardrails, memory, evaluations and connected tools may also process content outside the application.
The request arrives
Studio authenticates users with Cognito-backed sessions and checks workspace access in the application. The mobile app uses an authenticated session. Integration API keys are stored hashed; the plaintext is shown when a key is created.
The question is screened
Configured input guardrails check a request before the agent answers. Guardrails may make model calls, so a declined request may still send input to a model provider.
The library is navigated
The agent walks an index of your documents and opens pages by name. There is no embedding step and no vector store, so your library is not copied into a second representation held somewhere else.
The model is called
Through Amazon Bedrock or OpenRouter, according to the configured model and credentials. A workspace may use its own OpenRouter account. Inputs, conversation context, selected images and retrieved documents can be included in the request.
Output guardrails check the answer
Configured output guardrails check the response. Some checks are deterministic; model-based checks and evaluations can make additional provider calls. Their outcomes are recorded with the run.
The response carries its own evidence
A decision — answered, declined or escalated — the pages it opened, the guardrails that ran, and the usage that will bill. The same number your invoice is built from.
Use your own model provider account.
Point a workspace at your own provider key and inference is billed to you directly, at whatever rate you have negotiated. More importantly for this page: the question text and the retrieved library content then travel under your contract with that provider, on your terms. OpenRouter requests still require no provider data collection and zero-data-retention endpoints. Studio’s own conversation and run records have separate retention.
We take no share of model spend in either arrangement. When you do not supply a key it is passed through at what it cost us, which is the only reason our own route exists.
Content and metadata can both contain personal information. Retention depends on the record and its purpose; the debugging window does not govern every record.
Messages, inputs, outputs and attached images may persist separately from debugging traces. Workflow retention varies by record and run state. Contact your workspace administrator or support@krazimo.com to request export or deletion.
This setting applies only to captured model-request bodies used for debugging. Setting it to zero does not prevent Studio from storing conversations, run inputs and outputs, or attached images.
Model, timing, token counts and cost per call. It is what the monitoring screens and the cost accounting read. It is metadata about a call, not the content of one.
It is the billing record and the only one. The number a turn is metered on is the number you are invoiced for, which is why there is no separate counter that could disagree with it.
Your documents support retrieval by your agents. Removing a source document does not automatically remove excerpts already recorded in conversations, run results or backups.
What we hold, and what we do not
Underway and not yet complete. We are not certified today, and we would rather say that plainly than describe it in a way that reads as though we were. If your process needs a report before you can proceed, tell us where you are in your evaluation and we will tell you honestly whether our timeline can meet it.
The service uses Cognito-backed authentication and workspace access checks. The database is not publicly reachable, secrets are held as encrypted parameters rather than in the image or the task definition, and tenant API keys are stored hashed so the plaintext exists only at the moment you create one.
Questions this page did not answer?
Security review questions get a written answer from someone who has read the code, not a questionnaire response assembled from a template.