Private AI inference. Hosted, owned, paid once.

An OpenAI-compatible API
that you run, not rent.

A self-hosted gateway in front of an open-source LLM. Drop-in for any tool that speaks the OpenAI chat-completions format — no per-token bills, no rate-limited keys, no third-party seeing your prompts.

For humans Chat → Open WebUI — talk to the model in a browser. chat.deleg8.it For the operator Admin → Manage tokens, models, and inspect request traffic. admin.deleg8.it For your apps API → OpenAI-compatible endpoint. Send any model name. api.deleg8.it/v1

Quick start

Any OpenAI SDK or HTTP client works. Point your base URL here, drop in your token from the admin panel:

# .env
AI_STACK_KEY=sk-...
AI_STACK_BASE=https://api.deleg8.it/v1
AI_STACK_MODEL=qwen3
# curl
curl -X POST https://api.deleg8.it/v1/chat/completions \
  -H "Authorization: Bearer $AI_STACK_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3","messages":[{"role":"user","content":"hello"}]}'

What's actually running

Qwen3 30B-A3B Instruct (July 2025) on a single NVIDIA L4 GPU, fronted by LiteLLM and tunneled through Cloudflare. Hardware lives in AWS eu-west-2. Any model name a client sends (gpt-4, claude-3-5-sonnet, llama…) is routed to qwen3 internally, so apps with hard-coded model strings keep working.