FAQ

Access

Is the inference API live? It is deployed and switched off while it is in early access. Safeguard turns it on for each workspace that asks.

How do I get access? Open a model at safeguard.sh/models, choose Use with Safeguard and leave your email address. Safeguard turns the API Platform on for your workspace. Then follow the Quickstart.

Why am I getting 403 gateway_disabled? The API Platform is not switched on for your workspace yet. During early access it is off unless Safeguard has turned it on for you.

Why is Create key disabled? For the same reason: keys can be created once the API Platform is on for your workspace. Auditor accounts cannot create keys at all, because their access is read-only.

Compatibility

Is it compatible with OpenAI's API? With the Chat Completions API, yes: point an OpenAI SDK at https://api.safeguard.sh/gpt/v1 with an sg-api- key. See SDKs and clients.

Can I send images or audio? Not in v1. A message part that is not text is refused with 400 unsupported_parameter.

Are embeddings, image generation or the Responses API available? No. The API serves POST /chat/completions, GET /models, GET /models/{id} and GET /key.

Can I ask for several choices with n? No. n must be 1.

Do tools and JSON mode work? tools, tool_choice, parallel_tool_calls and response_format are passed to the provider unchanged, so they work where the model supports them. See Chat completions.

Models

Which models can I call? The ones GET /models returns. A catalog model becomes callable only after it passes Safeguard's launch checks. See Models.

Can I call Griffin, Eagle, Lion or Zero? Not through the API yet. See Safeguard models.

What do Safeguard's own models cost? The catalog lists Griffin at $0.0001 per million input tokens and $0.0001 per million output tokens, Eagle at $0.00001 and $0.00001, and Lion at $0.000001 and $0.000001. Zero uses no model tokens and is listed at $0. They are billed at those rates once they can be called through the API.

Does Griffin come in sizes, such as S, M or L? No. Griffin, Eagle and Lion are one model each, with no Lite, S, M, L or XL variants. Their parameters and context windows are in Safeguard models.

Can I use my own model or my own provider key? Not yet. API keys call models on Safeguard's provider accounts. The gateway accepts your own models (custom:<id>) only from the Console Workbench, and the Workbench cannot pick them yet. See Bring your own model.

Pricing and credit

Do you add a markup? 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.

How do I pay? Requests spend prepaid credit. Where card top-up is available, a tenant admin or super admin adds credit in the Console and the card fee is passed on at cost. Otherwise Safeguard adds credit for you. See Pricing and billing.

Do you send invoices? The Console issues no invoices for inference credit. Your record is the credit ledger on Usage. See Credits and invoices.

What happens when credit runs out? Requests stop: they are refused with 402 insufficient_credit until credit is added.

Can I cap what we spend? Yes. A tenant admin or super admin sets a monthly limit for the workspace, and each key can have its own monthly limit. See Spend controls.

Why was a request refused with 402 when I still had credit? Before each request the gateway reserves the most it could cost, based on max_tokens. Lower max_tokens, or add credit. See Limits.

Data

Where do my prompts go? To the provider of the model you name. If you allow fallbacks, a failed request can also go to a listed fallback model of the same family. Never to the models Safeguard uses for its own platform features.

Do you store my prompts? Not by default. The request log keeps metadata such as model, tokens, cost and latency. A tenant admin or super admin can turn on Store prompts and outputs for the workspace, and can delete what is stored. See Privacy and data.

How long do you keep request logs? Retention is configurable but not yet enabled: until Safeguard sets the periods, the request log and stored prompts are kept until deleted. Stored prompts can be deleted at any time from Usage. See Privacy and data.

Do you train on my data? No. Safeguard never uses your prompts or outputs to train models. Each provider's own terms apply to the requests it receives, and the catalog summarises them for each model.

Can I choose the region a request runs in? No. The gateway calls each provider's standard API and has no region setting. See Regions and data residency.

Can I require zero data retention? Yes, per request with provider.zdr, or for the whole workspace in Gateway settings. Requests then go only to models offered with zero data retention. See Zero data retention.

Keys and limits

How quickly does revoking a key take effect? Within 30 seconds.

What is the rate limit? 60 requests per minute per key by default. You can set a key's own limit from 1 to 10,000. See Limits.

Can I call the API from a browser or a mobile app? Don't. Anyone who can read the app can read the key. Call the API from your server.

Streaming

If I cancel a stream, do I pay less? No. The provider keeps generating, and the gateway reads the answer to the end to bill the provider's token counts. Use max_tokens to limit a request's cost.

MCP

Is the MCP server metered? No. MCP tool calls are unlimited on every plan, Free included, and never metered. See Inference API or MCP server.

Deployment and buying

Can we run the API Platform in our own environment? The Safeguard platform runs in public cloud, private cloud (VPC), on-prem and air-gapped deployments. The API Platform is in early access in all of them, and in an air-gapped deployment it has no model it can serve yet. See Deployment options.

Can we buy through our cloud marketplace? Not yet for the API Platform. See Procurement.

Help

Where can I see whether Safeguard is up? At status.safeguard.sh.

How do I get help? Use Help & Feedback in the web app, or email hi@safeguard.sh. See Status and support.