API reference

This page documents the supported inference endpoints for the current build environment. The routes below are the inference surface; there is no Responses API, no Anthropic Messages API, and no embeddings endpoint on this base URL, so do not point SDK features at it that assume them.

Authentication

Every request carries your key as an HTTP bearer credential:

Authorization: Bearer $TERALOR_API_KEY

(inject the variable from your secret manager — never a literal key). Requests with a missing or malformed Authorization header, or an unknown key, are rejected before any model work happens. A recognized key that is revoked, expired, or whose account/entitlement state refuses admission gets a 403 instead — Troubleshooting explains how to tell the two apart. The key is a secret — never in URLs, source control, browser code, or support messages.

GET /v1/models

Discovery. Requires auth and an empty body (a GET with a body is a 400). Returns the effective catalog — the intersection of the configured models and the currently healthy live capability subset, in canonical order — as an OpenAI-style list. Each entry carries at least:

Read the values from your response — this page deliberately does not print sample numbers, because any static numbers here could be mistaken for current limits. Treat discovery as authoritative for limits, capabilities, and the policy revision you are being served under. If a model is configured but its provider is unhealthy, it is simply absent from this list. Note also that discovery success reflects the state at that moment; a later completion can still be refused by budget, rate limiting, or runtime state.

POST /v1/chat/completions

Chat Completions for the Flash model. The Content-Type media type must be application/json (a charset parameter, e.g. application/json; charset=utf-8, is accepted); the body must be a JSON object, and the body size is bounded (oversized requests are rejected before parsing).

Allowed fields — every other top-level field is rejected with a 400:

Example (non-streaming):

curl -s https://api.teralor.com/v1/chat/completions \
-H "Authorization: Bearer $TERALOR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "model": "Flash",
  "messages": [{"role": "user", "content": "Summarize this changelog in two bullets."}],
  "max_tokens": 512,
  "stream": false
}'

Responses are standard OpenAI-style chat completions (an id, choices, usage when available). With "stream": true the same route streams chunks instead; intermediate transport errors simply end the stream — see Troubleshooting for the replay rules.

Tools. When tools is advertised and supplied, the assistant may return tool-call requests in its message. Your harness executes them under its own authority and sends role: "tool" results back in a follow-up request. The model never executes anything itself.

Context admission. Rendered input plus reserved max_tokens must fit the advertised context_length, and input must respect max_input_tokens; violations are 400s before dispatch — see Use Flash well for the budgeting math.

Errors

The API returns JSON errors in the shape {"error": {"message", "type", "param", "code"}} with an HTTP status matching the category. type is invalid_request_error for 4xx and api_error for 5xx. Rate/concurrency headers (retry-after, x-ratelimit-*) are forwarded when the platform supplies them. Edge infrastructure (a proxy or load balancer) can occasionally return its own HTML error page instead — treat non-JSON error bodies as an infrastructure signal, not an API response. The status narrows the category; the response message plus your account state identify the exact cause.

For what to do about each of these, see Troubleshooting.