THE ORRERY DEVELOPER PLATFORM
Generate with hosted models. Decide with Sentinel.
Build both into your application through one API platform.
OpenAI-compatible chat · Structured decisions · Usage controls
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.ephemerent.com/v1",
api_key=os.environ["ORRERY_API_KEY"],
)
reply = client.chat.completions.create(
model="doubleword-deepseek-v4-flash",
messages=[{"role": "user", "content": "Hello"}],
)Chat, code and tool calls
Score the available actions
Track requests and credits
MODELS & ROUTES
doubleword-deepseek-v4-flash · doubleword-glm-5-3-flashLanguage and code generation for your applications with hosted models (DeepSeek, GLM and others, served via Doubleword through Orrery Relay). Use familiar chat completions, streaming and tool-call formats.
sentinel-1An energy-based CLM for choosing between actions. Send context and candidate options; receive a structured decision.
orrery/fast · orrery/balancedUse a route alias to let Orrery select the serving model. Inspect response headers to see the provider and any fallback.
Model and route access depends on your account. Query GET /v1/models for the catalogue. A route name is not a guarantee of answer quality. Arbiter: available on request. Contact kt@ephemerent.com.
QUICKSTART
Choose a task. Copy an example.
Keep your API key on the server.
Loading example…API REFERENCE
https://api.ephemerent.com/v1Authorization: Bearer <key>/chat/completionsGenerate a responseSend model, messages and max_tokens (up to 8,192; the default). Set stream: true for incremental output. Other OpenAI fields outside the supported set, such as max_completion_tokens, seed and n, are ignored or rejected. Tools use the OpenAI function-calling schema; supported behavior depends on the serving model.
Returns choices and token usage. Inspect x-orrery-provider, x-orrery-route-alias and, when present, x-orrery-fallback-product. Optional Idempotency-Key supports replaying non-streaming requests.
/sentinel/decideChoose an actionRequired: state and question. Add options, a map of 2–16 names to descriptions, for multiple choice. Without options, the request is yes/no.
Named choices return answer, confidence and per-option p. Yes/no returns answer and p_yes. Scores compare the supplied options; verify important decisions against evidence. Sentinel is not a chat model.
/sentinel/rankRank candidatesSend context and 2–32 candidates, with an optional question. The response contains ranked pairs of candidate and score, best first.
/modelsDiscover models and routesList the catalogue using your API key. Use chat models with chat completions and sentinel-1 with its dedicated endpoints.
/usageCheck your creditsInspect monthly pools, used and remaining credits, and your purchased balance. The console provides the same account view.
/sentinel/statusCheck Sentinel availabilityCheck availability with your API key. If a decision endpoint returns 503, respect Retry-After and use your application's fallback behavior.
USAGE & CONTROLS
Hosted models are metered at the serving model's provider list rate, drawn from your plan's monthly pool, then from purchased credits. Fallback answers are metered at the fallback model's rate. Review your balance, current plan and credit packs in the console before sending paid requests.
View usage & credits →Create and revoke keys, restrict their routes, and review request receipts. Pro: 30 requests and 100k tokens a minute, 1 key. Max: 120 requests, 500k tokens, 5 keys. Ultra: 600 requests, 2M tokens, 20 keys. Context length depends on the serving model; requests are capped at 256 KB and 64 messages.
Manage API keys →Review your account's zero data retention setting for generation. Sentinel request and response bodies are not retained for training; usage metadata is recorded.
Read the data policy →BUILD SOMETHING THAT MOVES FORWARD