Skip to main content

​
NSR as a Model Provider

The NSR server speaks three wire protocols besides its native /v1/decisions surface, so existing tools can point at it unchanged. Every completion then runs through the neuro-symbolic pipeline — GSS grounding, policy rules, safety gates, refuse-by-default — instead of a bare LLM. All three support streaming (SSE), tools, and both auth styles — x-api-key: <key> or Authorization: Bearer <key>.

​
The model

The model is nazarae-1, with three mode variants: Legacy nsr-v2-* ids remain served as aliases for existing configs. GET /v1/models lists the aliases. Unknown model names are accepted and mapped by heuristic — *safe* → safe, *fast* or *mini* → fast, anything else → accurate — so a config naming another vendor’s model still works.
Set BASE to your deployment (e.g. https://api.stateset.ai, or http://127.0.0.1:8080 locally) and MODEL_API_KEY to your NSR API key in the examples below.

​
Claude Code

Claude Code streams from /v1/messages and preflights token counts via /v1/messages/count_tokens; both are served.
The Messages surface also accepts NSR extensions — facts, rules, and hydrate_org_context — in the request body, so a turn can be policy-grounded without leaving the Anthropic wire format.

​
Codex

~/.codex/config.toml:
wire_api = "responses" is required — recent Codex versions dropped wire_api = "chat". The chat wire is still served for other OpenAI-compatible clients.

​
OpenCode

​
Any OpenAI or Anthropic SDK

Point the SDK’s base_url at your NSR deployment and use any NSR API key. Errors come back in each provider’s native error envelope, so existing error handling keeps working.

​
What changes

The wire format is identical, but the behavior is not. Because every completion goes through the neuro-symbolic pipeline, the model will refuse rather than guess when it cannot ground an answer. Expect refusals where a bare LLM would have produced confident text — that is the intended behavior, not a failure.