API

One command scans the machine, shows every supported client with a checkbox, and writes the config files for the ones you tick. Everything lands in each client's own config file — no environment variables are touched unless you pick the shell target.

Linux and macOS:

curl -fsSL https://universalvoronez.tech/install.sh | sh

Windows, in PowerShell:

irm https://universalvoronez.tech/install.ps1 | iex

Non-interactive, for scripts and containers:

curl -fsSL https://universalvoronez.tech/install.sh | sh -s -- --key nrg-... --all

Remove it again — every change is reverted from the config files it touched:

curl -fsSL https://universalvoronez.tech/install.sh | sh -s -- --uninstall

Every file is backed up next to itself as *.nergate-backup before it is touched, and only the nergate entries are rewritten — your other providers, models and settings stay exactly as they were.

Issue a key in the console. It is shown once, starts with nrg-, and goes in the Authorization header.

curl https://universalvoronez.tech/openai/v1/chat/completions \
  -H "Authorization: Bearer nrg-..." \
  -H "content-type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.2",
    "messages": [{"role": "user", "content": "hello"}]
  }'

Python, official OpenAI client:

from openai import OpenAI

client = OpenAI(base_url="https://universalvoronez.tech/openai/v1", api_key="nrg-...")

client.chat.completions.create(
    model="zai-org/GLM-5.2",
    messages=[{"role": "user", "content": "hello"}],
)

opencode, in ~/.config/opencode/opencode.jsonc:

{
  "provider": {
    "nergate": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Nergate",
      "options": {
        "baseURL": "https://universalvoronez.tech/openai/v1",
        "apiKey": "{env:NERGATE_API_KEY}"
      },
      "models": {
        "zai-org/GLM-5.2": { "name": "GLM 5.2" }
      }
    }
  },
  "agent": {
    "nergate-max": { "options": { "reasoningEffort": "max" } }
  }
}

Anthropic SDKs need only the base — they append the rest:

from anthropic import Anthropic

client = Anthropic(base_url="https://universalvoronez.tech/anthropic", api_key="nrg-...")

client.messages.create(
    model="zai-org/GLM-5.2",
    max_tokens=256,
    messages=[{"role": "user", "content": "hello"}],
)

Bodies are forwarded as they came, apart from what is described below, so provider parameters work as documented upstream — stream, tools, reasoning_effort, temperature, max_tokens.

Streaming is a normal SSE stream. Nergate adds stream_options.include_usage so the final chunk carries token counts, which is also how your spend gets metered.

Three ways to set one. Each overrides the one before it.

On the key

Applies to every call made with that key. Set it in the console, or over the API:

curl -X PATCH https://universalvoronez.tech/api/me/keys/<key-id> \
  -H "content-type: application/json" -b cookies.txt \
  -d '{
    "systemPrompt": "You are a terse code reviewer.",
    "systemPromptMode": "replace",
    "defaultModel": "zai-org/GLM-5.2"
  }'

Per call, with headers

X-Nergate-System-Prompt: Answer in one sentence.
X-Nergate-System-Mode: replace

In the body

A plain {"role": "system"} message works as usual, and is what the two settings above act on.

Nothing is applied unless you ask for it: a model you call plain gets no reasoning_effort at all, and behaves as its own default. To pick a level, append it to the model id — the gateway strips the suffix and passes the effort up:

zai-org/GLM-5.2          # nothing added
zai-org/GLM-5.2-low      # fast and cheap
zai-org/GLM-5.2-max      # thinks hardest

Not every family accepts every level. GPT-OSS takes only low/medium/high, while GLM, DeepSeek and Kimi also take xhigh and max. Asking for a level a model does not support is a 400 unsupported_effort listing what it does accept — the gateway never silently substitutes another level. GET /openai/v1/models lists the variants each model has.

Four places can set it, later wins:

In opencode the level is an agent, not a model: the installer writes each model once and adds nergate-low…nergate-max alongside. Switch with opencode --agent nergate-max — the agent's effort applies to whichever model is selected.

X-Nergate-Effort: max

Models without reasoning — Llama, Gemma, Granite, Phi, Mellum — ignore all of this and have no variants.

These use a session cookie from /api/auth/login, not your API key — they are what the console itself calls.

GET /api/pricing is public and lists every model with your price per 1M tokens and the upstream list price. GET /openai/v1/models needs a key and returns the same rates in OpenAI shape, filtered to what your key may call.

Billing is prompt_tokens x input rate + completion_tokens x output rate, charged against your quota. Spend and remaining balance are on the console, and the full board is on the pricing page — 26 models.

Errors come back in OpenAI shape: {"error": {"message", "type", "code"}}.

A failing upstream account is retried on another one automatically, so pool_exhausted means several accounts failed in a row for the same call.