GetCortex27

OpenAI-compatible inference for open-weight agentic models. Operated by Minimorsel Enterprises Pvt. Ltd., Gurugram, India. Served from EU datacenters.

Models

Pricing and capacity are published machine-readably at /v1/models.

The conformance contract

Every endpoint we operate is held to the following, and we measure it on our own endpoints continuously:

1. Structured or nothing. A tool call is returned as tool_calls, or the request fails with an error. Raw tool-call markup never appears in content.
2. No empty successes. Every 200 carries content or tool calls, a finish_reason, and a usage object that includes reasoning tokens — streamed or not.
3. Parameters are honoured or rejected. A parameter we list as supported is enforced. One we do not support returns 400; it is never silently dropped.
4. Honest capacity. Above declared capacity we return 429 immediately, keyed by account, with Retry-After. We do not queue, and we do not degrade a response to stay under load.
5. Measured parity. We publish our own agentic-benchmark results for each endpoint, run on the production endpoint, with the harness.
6. Keep-alives. Long generations send SSE comment keep-alives so clients and routers never mistake thinking for a stall.

API

Base URL: https://api.getcortex.us/v1 (DNS pending). Endpoints: /v1/chat/completions (streaming and non-streaming), /v1/models, /health.

Data handling

Zero data retention: prompts and completions are not stored, not logged and never used for training. We keep request metadata only (timestamps, token counts, model version, status, latency) for billing and reliability. Full terms in the privacy policy.