oriora-c2 0.1.2

Oriora c2 — a local proxy that plugs Oriora's model-routing decision into any OpenAI-style agent. Your vendor key and prompts stay on your machine; only the routing decision crosses.

oriora-c2

Use Oriora's model-routing decision in any OpenAI-style agent — without your key or prompts ever leaving your machine.

oriora-c2 runs a tiny local proxy on 127.0.0.1. Your agent points at it; before each call it asks Oriora's /api/select "which model is best for this task?", then dispatches the call directly to the vendor on your own key. Only the routing decision (task type + your candidate models) crosses to Oriora — never the key, the prompt, or the response.

This is for off-the-shelf agents (Cursor, Aider, Continue, the raw openai SDK, LangChain ChatOpenAI, …) that can't easily insert a "ask Oriora first" step. Writing your own code? You don't need this — call /api/select directly (or pip install oriora and use model_select()).

Install

pip install oriora-c2     # Python 3.10–3.13

Python 3.14 is not supported yet — a dependency (litellm's pinned orjson) has no 3.14 build. If python3 --version says 3.14, create the venv with python3.13 -m venv.

Run

oriora-c2 init                       # scaffolds config.yaml + .env.oriora-c2.example
# set your keys:
export ORIORA_API_KEY=sk_oriora_...  # the decision call only
export DEEPSEEK_API_KEY=...          # your own vendor keys (the actual call runs on these, locally)
export MINIMAX_API_KEY=...
oriora-c2 serve                      # local proxy on http://127.0.0.1:4000

Point any OpenAI client at it:

from openai import OpenAI
c = OpenAI(base_url="http://127.0.0.1:4000/v1", api_key="anything")
c.chat.completions.create(model="oriora-auto", messages=[{"role":"user","content":"…"}])
# task type is auto-detected locally (free); or force it: model="oriora-auto:coding" (or :coding_hard)

Reasoning models behave vendor-native through the proxy. The proxy never touches the response, so models like deepseek-v4-pro return their reasoning_content raw — and a tiny max_tokens budget can be consumed by reasoning before any answer text appears. Budget max_tokens generously (or pick non-reasoning candidates) for short-answer use.

How it works

1. Agent → http://127.0.0.1:4000 (model="oriora-auto"). Prompt never leaves your box.
2. The pre-call hook classifies the task locally (free regex rules + overshoot-biased difficulty escalation to *_hard, no LLM) → calls POST /api/select {task_type, models} — the one Oriora touch ($0.001/decision).
3. It rewrites data["model"] to the recommended model.
4. LiteLLM dispatches direct to the vendor on your local key; the stream flows vendor → you.

Privacy / c2 invariant: the proxy is customer-hosted (127.0.0.1). Only {task_type, model candidates} reach Oriora. If a third party ever hosted this, it would no longer be c2.

Fail-open: if /api/select is slow (>ORIORA_SELECT_TIMEOUT_S, default 2.5s) or down, the hook falls back to ORIORA_FALLBACK_MODEL so your agent is never blocked.

Configuration (env)

Var	Purpose
ORIORA_API_KEY	Oriora key for the decision call (required)
DEEPSEEK_API_KEY, MINIMAX_API_KEY, …	your own vendor keys (the call runs on these)
ORIORA_CANDIDATES	comma-sep catalog ids you hold keys for (sent to /api/select)
ORIORA_FALLBACK_MODEL	model used if the decision call fails (default deepseek-v4-flash)
ORIORA_SELECT_TIMEOUT_S	decision-call budget before fail-open (default 2.5)

Add a vendor = add its key + a model_list entry in config.yaml + its catalog id to ORIORA_CANDIDATES. v1 ships configured for DeepSeek + MiniMax (OpenAI-format).

MIT © Orioralabs OÜ · https://orioralabs.com