UNIFIED MIXERS

the workgroup model gateway

One binary that owns model traffic on machines you control. It holds your servers — vLLM, llama.cpp, Ollama, LM Studio, OpenRouter — and hands every client (Copilot BYOK, opencode, codex, DSH, curl) one clean API. System prompts are rewritten server-side, broken tool calls repaired, flaky exchanges retried, every byte captured, the ledger kept here instead of in every client.

PRE-M1 · BUILDING IN PUBLIC

Where the gaps are

Our claim, as of October 2026, from a review of roughly 30 products — per-claim findings published with citations in the research report. If something shipped these four together and we missed it, tell us.

  1. Declarative per-model system-prompt injection for arbitrary OpenAI clients — personalities as rule files, applied upstream, invisible to the client.
  2. Named per-model parameter presets surfaced in the catalog — qwen3.8:Think, the variant grammar OpenRouter clients already speak.
  3. First-class local capture of upstream bytes, before and after rewriting, with replay. No external observability stack.
  4. The single-owner, loopback form factor — one binary, no Postgres, no Redis, no user table. Workgroups share it through hand-minted tokens with caps and attribution.

License

You, your lab, your lecturesFree — every feature. Non-production use, plus personal, educational, and evaluation production use.
Your company, in productionOne flat annual fee per organization. No sales call; the invoice beats the audit.
Everyone, in four yearsEvery version converts to the MIT License. Source-available today, open source later — by design.
github.com/unifiedmixers/gateway read the plan