Security and data flows
What stays with you, and what kibbu receives.
With kibbu, your AI requests run on machines your company owns. Hosted, kibbu's control plane routes each request to one of those machines, so the request passes through it, and it keeps only metadata such as the model, token counts and timings. With kibbu On-Prem you run the control plane yourself, and nothing reaches kibbu.
Cloud providers are never used unless you add your own provider key and your policy allows it.
By deployment
Where your requests go.
Inference runs on your machines in all three. The difference is who runs the control plane, and what it can reach.
| Hosted | kibbu On-Prem, connected | kibbu On-Prem, air-gapped | |
|---|---|---|---|
| Control plane | Run by kibbu. | Run by you, in your datacenter or your own cloud account. | Run by you, with no route to kibbu. |
| Inference | On your enrolled machines. | On your enrolled machines. | On your enrolled machines. |
| Request path | Your app calls the kibbu API. The hosted control plane reads each request to route it and passes it to one of your machines over an encrypted mesh, then streams the answer back. It does not write prompt or response text to its request logs. | Requests go from your app to your control plane to your machines. None of it reaches kibbu. | Requests stay inside your environment from end to end. |
| What kibbu receives | Request metadata (model, token counts, timings, status, which machine served it), machine health, agent warning logs and crash reports. Request text only where you use a feature that stores it (see stored content). | Nothing about your usage. The control plane sends kibbu no version, usage or machine count. It only downloads releases and model files from kibbu's registry, with your download credential. | Nothing. |
| Cloud providers | Only if you add your own provider key and your key policy allows cloud. | Only if you add your own provider key and your key policy allows cloud. | Not used, unless you deliberately configure a provider, which the egress check then reports. |
| Network | Machines need outbound HTTPS to kibbu and outbound UDP for the encrypted mesh. | Outbound to kibbu's registry for releases and models, plus the services you configure. | None outside your environment. Releases and model files arrive as a signed offline bundle or through your own registry. |
The On-Prem page covers what running the control plane yourself involves.
On your machines
What stays on the machines that serve.
Any enrolled machine can serve any request its key policy allows. Use machine groups to keep sensitive work on the machines you choose.
- Model weights, in a local cache (20 GB by default) that the agent manages.
- While a request runs, its prompt and answer exist in the machine's memory.
- The inference engine can keep a prompt cache on disk to speed up repeated work.
- Your traffic never trains any model.
What kibbu receives
What the control plane keeps.
This is what a hosted control plane receives. With kibbu On-Prem, your own control plane receives it, and kibbu receives nothing.
| What | Contents |
|---|---|
| Machine health | CPU and memory use, power source, battery level and trend, thermal pressure, GPU memory and use, network rates, how long the machine has been idle and whether the screen is locked, which models it has cached, its queue, and the agent's version. It does not include prompts, responses or files. |
| Request log | For each request: organization, model, API key and key policy, which machine served it, whether it ran on your fleet or on a cloud provider, token counts, timings, status and error class. Not the prompt or the response. |
| Admin audit log | Who did what in the admin console: sign-ins, API key issuance and revocation, model enablement, machine approval and removal. |
| Agent logs | Agent warnings and errors, and crash reports with the last lines of the agent's own output. Not request content. |
| Stored content | Only where you use a feature that keeps it, listed below. |
Where request text is stored
- /v1/responses keeps a response for 30 days so it can be retrieved or continued, unless the request sets store: false.
- Files uploaded through /v1/files and prompts saved through /v1/prompts are kept until you delete them.
- Sessions saved in the console playground.
- The idempotency cache keeps a response for 24 hours when a client sends an Idempotency-Key, so a retried request gets the same answer.
Retention
How long it is kept.
Defaults. An admin can change retention in the console.
| Record | Kept for |
|---|---|
| Request metadata | 90 days |
| Admin audit log | 365 days |
| Signed request receipts | 365 days |
| Machine telemetry | 30 days |
| Crash reports | 30 days |
| Stored /v1/responses content | 30 days, or not stored with store: false |
| Idempotency replay cache | 24 hours, only when the client sends an Idempotency-Key |
Access and policy
Who can do what, and where work may run.
The admin console has two roles: admins manage the organization, and members can sign in and use the apps an admin has connected. Sign-in goes through your identity provider.
Every API key can carry a key policy, and the organization has a default policy for keys without their own. A policy can set:
| Rule | What it controls |
|---|---|
| Cloud | Allowed, denied, or limited to named providers. |
| Models | An allowlist of the models a key may call. |
| Machine groups | Which groups of machines may serve the key, such as your servers only. |
| Countries | Which countries the serving machines may be in. |
| When the fleet is busy | Fall back if allowed, return an error, or wait up to 60 seconds for a machine. |
Cloud providers
Cloud only with your key and your permission.
A local-only setup is one where no provider key is added, or every policy denies cloud. Requests then never leave your fleet.
kibbu does not run a cloud of its own. A request can only reach a cloud provider if your organization has added that provider's API key, the model allows cloud, and the key's policy does not deny it. With cloud denied, a request that no machine can serve returns an error (HTTP 503 with a Retry-After header) rather than leaving your fleet.
Supported providers: OpenAI, Anthropic, Google, Mistral, Together, Groq, Amazon Bedrock, Google Vertex AI, Azure OpenAI, or a custom OpenAI-compatible endpoint, using your own keys.
A request served by a cloud provider is subject to that provider's terms, and the request log records that it ran there.
Compliance
Certification status.
Stated exactly as it is.
SOC 2 Type II and ISO 27001 are in progress, not yet certified. kibbu holds no other certification. Ask hello@kibbu.io for the current security questionnaire.
The network traffic between the control plane and your machines runs over an encrypted mesh. Report a vulnerability to security@kibbu.io.
Keep reading
Start with one workflow.
Tell us about the work you want to keep private, and we will scope a pilot on your own machines.