ogentic-router

Privacy-aware LLM routing. Sensitive content stays local; redacted content may go to cloud. Pairs with ogentic-shield for classification and ogentic-audit (coming) for tamper-evident evidence.

⚠️ Pre-alpha. The repo is scaffolded; v0.1 build is in flight. Track progress on the Linear project.

Thesis

Every "AI control plane" today (Merge Gateway, OpenRouter, Helicone, Portkey, LiteLLM proxy) routes on cost / latency / quality. None routes on content sensitivity. Result: the routing decision is made after the content has already left the user's machine — which structurally disqualifies these products for regulated content (legal privilege, PHI, MNPI).

ogentic-router inverts the architecture: the routing decision happens on the user's device, before the content leaves. The Shield classifier decides sensitivity → the router picks a backend → only the cleared payload crosses the network. Pairs with ogentic-audit to produce a court-defensible record of every routing decision.

This is the v0.1 promise. The library, not the service.

Architecture

┌──────────────────────────┐
│  Application code        │   uses standard OpenAI / Anthropic SDK
└──────────────┬───────────┘
               │ endpoint swap (drop-in)
               ▼
┌──────────────────────────┐
│  ogentic-router (local)  │
│  1. Shield.analyze       │  classify input
│  2. Policy engine        │  sensitivity + intent → backend choice
│  3. Adapter              │  call the picked backend
│  4. Audit emit           │  ogentic-audit decision row
└──────────────┬───────────┘
               │
   ┌───────────┼─────────────────┐
   ▼           ▼                 ▼
Local LLM   Cloud (cleared)   Cloud (redacted)

Key invariant: Shield + Policy + Audit run in-process on the user's machine. The router is a library, not a hosted service. There is no Ogentic infrastructure that ever sees user content.

How it differs from SaaS gateways

v0.1 scope

• Pluggable Shield classifier (default: ogentic-shield).
• Pluggable backend adapters: OpenAI, Anthropic, Ollama, llama.cpp.
• Declarative policy: if sensitivity ≥ N OR group ∈ {PRIVILEGE, PHI, MNPI} → local; else → cloud (with optional Shield.redact prefilter).
• Endpoint-swap server: localhost:NNNN/v1/chat/completions, drop-in for existing OpenAI/Anthropic SDKs.
• Decision logging into ogentic-audit.
• Streaming + non-streaming.
• Python library + CLI + MCP tool surface.

Install (pre-alpha)

pip install ogentic-router  # not yet published; coming with v0.1

For now, install from source:

git clone https://github.com/OgenticAI/ogentic-router
cd ogentic-router
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"

Status

v0.1 milestone target: 2026-08-31. See the Linear project for tickets and progress.

License

Apache-2.0. See LICENSE.

Ecosystem