Continue.dev + Gonka AI - AI for VS Code/JetBrains

Continue uses ~/.continue/config.yaml (the old config.json format is deprecated). Open it via the command palette: Ctrl+Shift+P (on macOS — Cmd+Shift+P) → "Continue: Open config.yaml".

Add JoinGonka Gateway as a provider:

# ~/.continue/config.yaml
name: My Config
version: 1.0.0
schema: v1
models:
  - name: Gonka DeepSeek V4 Flash
    provider: openai
    model: deepseek-ai/DeepSeek-V4-Flash-0731
    apiBase: https://gate.joingonka.ai/v1
    apiKey: jg-your-key
    roles: [chat, edit, apply]
    capabilities: [tool_use]
    defaultCompletionOptions:
      contextLength: 380000
      maxTokens: 32768
  - name: Gonka GLM-5.3 Flash
    provider: openai
    model: zai-org/GLM-5.3-Flash
    apiBase: https://gate.joingonka.ai/v1
    apiKey: jg-your-key
    roles: [chat, edit, apply]
    capabilities: [tool_use]
    defaultCompletionOptions:
      contextLength: 390000
      maxTokens: 8192
  - name: Gonka MiniMax M2.7
    provider: openai
    model: MiniMaxAI/MiniMax-M2.7
    apiBase: https://gate.joingonka.ai/v1
    apiKey: jg-your-key
    roles: [chat, edit, apply]
    capabilities: [tool_use]
    defaultCompletionOptions:
      contextLength: 200000
      maxTokens: 8192
  - name: Gonka Autocomplete
    provider: openai
    model: deepseek-ai/DeepSeek-V4-Flash-0731
    apiBase: https://gate.joingonka.ai/v1
    apiKey: jg-your-key
    roles: [autocomplete]

The header (name, version, schema) is needed if the file doesn't exist yet; if the file already exists, just add the entries to the models list. The defaultCompletionOptions block declares the model's real context window and response ceiling — without it, Continue will substitute its own values for an unrecognized name.

Continue shows models with the chat, edit, apply roles as a list — switch right in the chat: DeepSeek V4 Flash holds a 380K context and responses up to 32768 tokens (large files and long edits), GLM-5.3 Flash is a reasoning model for tangled logic (it reasons before answering, so it takes longer to respond), MiniMax M2.7 is faster on short responses. Autocomplete is powered by DeepSeek V4 Flash, and that's not a matter of taste: Continue picks the suggestion format based on the model NAME, and among the network's models only DeepSeek understands it. MiniMax responds to the same markup with reasoning instead of code, and GLM-5.3 Flash as a reasoning model isn't meant for autocomplete — the suggestions will be garbage.

Autocomplete over a decentralized network is slower than Copilot's: about two seconds per suggestion versus a fraction of a second. Chat, edits and agent mode are unaffected.

The capabilities: [tool_use] line is required for agent mode. Continue enables tools only for models on its list (gpt-*, o*, codex*) — Gonka's models aren't there, and without this line the agent is left without tools: it can't read files, can't edit code, and the chat can't see the project. The gateway itself does support tool-calling.

Project indexing (@codebase) requires an embedding model — Gonka doesn't have any on the network yet. Add context via @file and @folder, or work in agent mode, which reads files with tools.

To avoid keeping your key in plain sight, Continue supports substituting apiKey: ${{ secrets.GONKA_API_KEY }} with a variable in ~/.continue/.env.

Check: open any file with code and start typing — autocomplete should work. Or open the Continue chat (Ctrl+L) and ask a question.

Important: make sure apiBase includes /v1 at the end of the URL. Without it, requests will return 404.