Private AI Endpoints

A private endpoint lets your agents call your own model endpoint instead of a model SchemaBounce runs for you. You keep your provider contract, your data terms and your bill. SchemaBounce stores the key, adds it to each call and never shows it to an agent.

Private endpoints are included on the Team, Scale and Enterprise plans.

What you can connect

Any endpoint that speaks the OpenAI Chat Completions API over https. Four presets fill in the URL shape for you:

Endpoint	URL you enter	Key	Model names
Azure OpenAI	https://your-resource.openai.azure.com/openai/v1	Azure OpenAI key	Your deployment names (required)
Amazon Bedrock	https://bedrock-mantle.us-east-1.api.aws/v1	Bedrock API key	Optional; the endpoint lists its models
Databricks Model Serving	https://your-workspace.cloud.databricks.com/serving-endpoints	Databricks token	Your serving endpoint names (required)
OpenAI-compatible server	The base URL in front of /chat/completions	That server's key	Optional; read from its /models

The OpenAI-compatible preset covers vLLM, LiteLLM, NVIDIA NIM, Ollama's /v1 API, Text Generation Inference and company AI gateways.

The URL is everything before /chat/completions. For Azure that means the v1 API form ending in /openai/v1, with no api-version parameter.

Add an endpoint

1. Open Models & providers, then the Providers tab. The inference mode must be Bring your own keys.
2. Under Private endpoints, choose Add endpoint.
3. Pick the endpoint type, give it a name, and enter the URL and key.
4. For Azure OpenAI and Databricks, list the deployment or serving endpoint names agents may call, one per line.
5. Choose who can use it: everyone in the workspace (needs the providers:manage permission) or only you.
6. Choose Test connection, then Save endpoint.

SchemaBounce checks the key before saving. It lists the endpoint's models, or, when the endpoint has no models list, sends one request of one token to the first model name you entered.

Use it in an agent

On the agent's Model settings, choose Private endpoint as the provider, pick the endpoint, then pick a model. The provider choice appears once your workspace has at least one private endpoint.

How the key and the URL are protected

• The key stays out of the agent. It is stored encrypted with your workspace's own key. The agent's runtime holds only a reference; the SchemaBounce gateway adds the key to each request.
• The key travels with its URL. The key and the URL are stored together. Changing the URL needs the key again, so a saved key cannot be pointed at a different host by anyone who does not already hold it.
• Public https only. The URL must be https on port 443. Private, loopback, link-local and cloud metadata addresses are refused when you save, and again on every connection, including a hostname that later resolves to one of them. Redirects are never followed.
• Every change is in the audit log, with the endpoint type and host. Keys and request bodies are never logged.

Endpoints reachable only from a private network or from an IP allowlist are not supported yet.

Billing

Your provider bills you for the tokens. SchemaBounce does not charge credits for a run on a private endpoint. Each run still has a cap: because SchemaBounce has no price for your model, the per-run cost cap becomes a token ceiling of 2,000,000 billable tokens by default, which you can change per agent.

Limits

• Ten private endpoints per workspace.
• Key-based authentication only. Microsoft Entra ID, AWS IAM roles and Google Vertex AI service accounts are not supported yet.
• Agents on a private endpoint cannot use the image viewing tool yet.

Common questions

Does SchemaBounce store my prompts on my endpoint's side?

No. Requests go to your endpoint under your provider's terms. What your provider keeps is set by your agreement with them, for example Azure OpenAI's data and abuse-monitoring settings for your resource.

Why does Azure OpenAI need deployment names?

Azure routes a request by the deployment name you created, not by the base model name. Its models list shows base models, so SchemaBounce offers only the deployment names you enter.

What happens to an endpoint if the workspace leaves the Team plan?

The endpoint stays saved, but agents cannot run on it until the workspace is back on Team or higher.

Can I use Claude through Bedrock or Vertex AI?

Not through a private endpoint yet. Bedrock's OpenAI-compatible endpoint serves open-weight models; Claude on Bedrock and Vertex AI need their own request formats, which are planned.