Safeguard Docs
API Platform (Console)

Changelog

What has changed in the API Platform (Console) and its inference API, newest first.

14 October 2026

Each item below is generally available; where your workspace does not have access yet, nothing changes.

  • Routing. Chat and completions requests can list more models in models, prefer, limit or exclude providers with provider.order, only and ignore, and sort candidates by price, latency or uptime with provider.sort. The X-Route header names the provider route that served the answer. See Routing.
  • Embeddings. POST /embeddings for embeddings models, billed per input token at the provider's list price with no markup. The text you send is never stored. See Embeddings.
  • Completions. POST /completions accepts the legacy single-prompt format for chat models. See Completions.
  • Console. A Models page lists every model with its provider, context length, prices per 1M tokens and measured latency and uptime; the Workbench adds Chat, Compare (two or three models side by side) and a code panel. See Workbench.
  • Your own provider keys. Workspace administrators can add a provider key per provider; requests for that provider then run on it, billed by the provider with no Safeguard fee. The key is sealed and never shown again. See Your own provider keys.
  • Usage statements. Usage lists every month with activity and downloads each as CSV. See Credits and invoices.
  • Key creation explains where prompts go. The prompt goes to the provider of the model you choose, per request, and Safeguard never uses API traffic to train models. See Privacy and data.

9 October 2026

8 October 2026, second update

  • One rule for which models can be called. A model is available only when the catalog lists it as available and it passes the launch checks; otherwise it shows request_access or coming_soon. GET /models?all=true, the Workbench and the gateway all use the same rule. Each model now carries runnable and, when it cannot be called, an unavailable_reason. Models not published in the catalog are never listed or served.
  • X-Generation-Id on every chat response. Streamed or not, the header carries the request id (gen-…), as does an error returned after the request was accepted. X-Request-Id is now only an internal id for Safeguard's audit trail. The Workbench shows each run's Request ID.
  • Keys by name, never by id. Keys are shown masked, as their start and end (sg-api-1a2b3c4d…9f8e). Usage and the request log name each key by its name and masked form, Workbench runs as Console workbench, and a key that no longer exists as Deleted key. Revoked keys stay listed under Revoked on API keys.
  • Key edits are recorded as auth.api.key.updated, naming only the fields that changed. A save that changes nothing records nothing.
  • Tenant admins and super admins manage the gateway settings, card top-up and stored prompts. Safeguard's billing service enforces it, and each change is recorded in the workspace's audit trail with the role it was made under.
  • The zero data retention setting is now called "Only use models whose provider offers zero data retention". Its behaviour is unchanged: other models are refused.
  • Times are sent in UTC, as ISO 8601 with Z, and the Console shows them in your browser's time zone.

8 October 2026

  • The inference API is deployed. It serves POST /chat/completions with streaming, GET /models, GET /models/{id} and GET /key at https://api.safeguard.sh/gpt/v1, in OpenAI's Chat Completions format. It stays off until Safeguard turns it on for a workspace.
  • API keys are issued by Safeguard's servers. A key is sg-api- and 64 hexadecimal characters, carries only the inference scope, is shown once and is stored as a hash. Each key can have a monthly spend limit and a requests-per-minute limit, and a revoke takes effect within 30 seconds.
  • Prepaid credit and spend limits. Each request reserves its maximum cost before the provider is called and is charged its actual cost afterwards, with no markup on third-party models. Requests stop at zero credit with 402. Monthly limits apply per workspace and per key. The request log keeps metadata only, and storing prompts is opt-in. Card top-up is built and switched off; until it is on, Safeguard adds credit.
  • Launch checks for every model. A model can be called only when the catalog lists it as available with prices and a successful end-to-end test call is recorded. For third-party providers, the provider's terms, its place on the Trust Center subprocessor list and Safeguard's legal documents are checked first.
  • One catalog for the website and the API. safeguard.sh/models and the gateway read the same model catalog. Safeguard Zero is listed with no model tokens, at $0 per million tokens.
  • The Console is connected to the gateway. Where the API Platform is on: keys are created, edited and revoked on the server; the Workbench streams through the gateway and shows each run's cost; Usage shows the gateway's metering, the request log, credit and the credit ledger; tenant admins manage the workspace's spend limit, prompt storage, fallbacks and zero data retention. Where it is off, Create key is disabled.
  • These docs moved to /docs/console, with pages for every part of the API. Links to /docs/api-platform redirect here.

5 October 2026

  • The Console was added to the web app as Safeguard's fifth product, first for workspaces Safeguard gave access: Dashboard, Search, Workbench, API keys, Usage and Settings, opened from the product switcher or with G then 5. At that stage the inference service behind it did not exist yet, so no request reached a model.

On this page