GitHub Copilot BYOK — own models instead of quotas

A Copilot subscription restricts premium requests with monthly limits; agentic sessions, refactoring, and multi-step tasks consume this quota fastest. BYOK shifts these calls to your provider: consumption is billed at your provider's rates and does not deduct from the Copilot premium quota. Keys are stored in the system keychain and cannot be read back by the interface.

BYOK is justified if you are hitting limits, want models with open weights, or want to pay significantly less for inference: the Gonka network offers the same open-source models hundreds of times cheaper than vendor retail prices — current rates are always available in the live gateway price list.

You'll need a jg-… key — it's issued when you register on the gateway, along with 3M free starter tokens, no card required. The key and model are set separately in each Copilot client — here are three ways.

VS Code — Custom Endpoint provider. This is the current path: the former "OpenAI Compatible" provider and the github.copilot.chat.customOAIModels setting are now deprecated.

"models": [
  {
    "id": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "name": "DeepSeek V4 Flash (Gonka)",
    "url": "https://gate.joingonka.ai/v1/chat/completions",
    "toolCalling": true,
    "vision": false,
    "maxInputTokens": 347232,
    "maxOutputTokens": 32768
  },
  {
    "id": "MiniMaxAI/MiniMax-M2.7",
    "name": "MiniMax M2.7 (Gonka)",
    "url": "https://gate.joingonka.ai/v1/chat/completions",
    "toolCalling": true,
    "vision": false,
    "maxInputTokens": 191808,
    "maxOutputTokens": 8192
  },
  {
    "id": "zai-org/GLM-5.3-Flash",
    "name": "GLM 5.3 Flash (Gonka)",
    "url": "https://gate.joingonka.ai/v1/chat/completions",
    "toolCalling": true,
    "vision": false,
    "maxInputTokens": 381808,
    "maxOutputTokens": 8192
  }
]

maxInputTokens is the model's context window minus the response ceiling: together with maxOutputTokens it defines the window. Without toolCalling: true, the model won't appear in the list when working with agents.

Copilot CLI — environment variables only. This provider has no settings file; values are set in the shell from which you launch copilot:

export COPILOT_PROVIDER_BASE_URL=https://gate.joingonka.ai/v1
export COPILOT_PROVIDER_TYPE=openai
export COPILOT_PROVIDER_API_KEY=jg-your-key
export COPILOT_MODEL=deepseek-ai/DeepSeek-V4-Flash-0731
export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=347232
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=32768
copilot

Only Copilot CLI reads the COPILOT_PROVIDER_* names, so this setup won't interfere with your other tools. The CLI requires tool calling and streaming from the model — every model in the network supports both.

JetBrains — GitHub Copilot plugin. Open Copilot Chat → model list → Manage Models, choose a provider with an OpenAI-compatible endpoint, click Add Models and enter the Base URL https://gate.joingonka.ai/v1, your jg-… key and the model ID. Click Save, check the model under your key and save again — it will now appear in the list alongside the built-in ones.

Tip. The npx @joingonka/setup --tool copilot-byok command prints all three sections with your key and the actual limits of the selected model already filled in — including a ready-to-paste "models" array for VS Code — and verifies with a live request that the gateway accepts the key and the model. It writes no files: each Copilot client keeps the key in its own storage.

All three network models support streaming and tool calling. deepseek-ai/DeepSeek-V4-Flash-0731 has one of the longest contexts in the network (380K tokens) and the largest output, up to 32768 tokens: use it for large repositories and long agent sessions. MiniMaxAI/MiniMax-M2.7 is universal, with an output up to 8192 tokens. zai-org/GLM-5.3-Flash is Z.ai's reasoning model with the longest context in the network (390K tokens): it reasons before answering, with an output up to 8192 tokens; part of the response budget goes to reasoning, so don't set the output limit too low in the settings.

BYOK covers agent sessions and chat; inline suggestions and semantic project search remain on Copilot models and require a GitHub account — this is a limitation of Copilot itself, not the provider. In Copilot Business and Enterprise, the local BYOK is managed by administrator policy ("Bring Your Own Language Model Key"): if the option is not in the interface, it is disabled at the organization level. Vision tasks do not work through the gateway: network models are text-only. If a tool sends requests in the Responses API format, the gateway will accept them: POST /v1/responses, streaming, and function tools work. We do not store dialogues, so store is ignored, and previous_response_id will return an error — pass the entire history in input.