API Platform (Console)
Safeguard's OpenAI-compatible inference API and its Console: one API key for models from several providers, billed at the provider's list price with no markup. In early access.
The API Platform is Safeguard's fifth product, alongside ESSCM, Portal, TPRM and OSM. It gives your applications one OpenAI-compatible API for model inference. You send chat completion requests to Safeguard with an API key, Safeguard routes each one to the model you named, and every request is metered against your workspace's prepaid credit.
The Console is the API Platform's interface in the Safeguard web app. It is where you create API keys, try models in the Workbench, follow usage and cost, and manage credit and spend limits.
Early access
The inference API is deployed but switched off while it is in early access. To use it, open a model at safeguard.sh/models and choose Use with Safeguard. Safeguard then turns the API Platform on for your workspace. Until it is on, Create key in the Console is disabled, and a request is refused with 403 gateway_disabled.
What you get
- One API, OpenAI's shape.
POST /chat/completionswith streaming, plusGET /models, at the base URLhttps://api.safeguard.sh/gpt/v1. The OpenAI SDKs work once you change the base URL and the key. See SDKs and clients. - Models from several providers in one catalog. The catalog at safeguard.sh/models lists third-party models and Safeguard's own. Through the API, a model can be called only after it passes Safeguard's launch checks. See Models.
- No markup on model usage. For third-party models you pay the provider's list price per token. Each response reports its own cost in
usage.cost. See Pricing and billing. - Spending you control. Prepaid credit with a hard stop at zero, a monthly spend limit for the workspace, and an optional monthly limit and requests-per-minute limit for each key. See Limits.
- Metadata-only logs by default. The request log keeps model, tokens, cost, latency and status. Storing prompts and outputs is off unless a tenant admin or super admin turns it on. See Privacy and data.
- Your routing rules. Zero data retention routing when you ask for it, and retries or fallbacks only when you allow them, within the same model family. See Reliability and fallbacks.
Get access
- Choose a model. Open safeguard.sh/models, pick a model and choose Use with Safeguard. The form asks only for your email address.
- Safeguard turns it on. Safeguard switches the API Platform on for your workspace. Requests spend prepaid credit; while card top-up is not available, Safeguard adds the credit to your workspace.
- Create a key and call the API. In the Console, open API keys and create a key. The Quickstart takes you from there to your first request.
The Console
When the API Platform is on for your workspace, API Platform (Console) appears in the product switcher, and pressing G then 5 opens it. Its sidebar has Dashboard, Search, Workbench, API keys, Usage and Settings. The Console guide describes each page.
Delivery options
| Option | What it is | Maturity |
|---|---|---|
| Safeguard API (hosted) | Safeguard runs the gateway. You call https://api.safeguard.sh/gpt/v1 with a key from the Console. | Early access |
| Public cloud | Safeguard deployed on the public cloud you choose. The platform is cloud-agnostic, so it is not tied to one provider. | GA |
| Private cloud (VPC) | A single-tenant deployment in a private network, isolated for one customer. | GA |
| On-prem | The full platform in your own datacenter. | GA |
| Air-gapped | The complete platform with no outbound connection. | GA |
| Your cloud marketplace | Buying the API Platform through your cloud provider's marketplace. | Not available yet |
The GA labels describe running the Safeguard platform in that environment. The API Platform itself is in early access in every option. In a deployment of your own, it can call a provider only when that deployment can reach the provider's API, so in an air-gapped deployment it has no model it can serve yet. Deployment options and Procurement have the details.
v1 limits
- Text only. A message with an image or audio part is refused with
400 unsupported_parameter. - One choice per request:
nmust be1or absent. - Chat completions and models only. There are no embeddings, images, audio or legacy completions endpoints.
- Your own models (
custom:<id>) are accepted only from the Console Workbench, never with an API key, and the Workbench cannot pick them yet. See Bring your own model. - Safeguard switches the API Platform on per workspace.