Skip to content

Support bring-your-own LLM (hosted OpenAI-compatible / Azure OpenAI) for language model integration #2034

Open

Description

@waldekmastykarz

Summary

Dev Proxy's language model integration is currently designed for local SLMs (default: Ollama llama3.2 at localhost:11434). As inference prices drop, using hosted LLMs is increasingly viable and yields better results (e.g. operation IDs/descriptions in OpenApiSpecGeneratorPlugin/TypeSpecGeneratorPlugin, mock responses in OpenAIMockResponsePlugin).

Current limitations

  • OpenAILanguageModelClient sends no Authorization/api-key header, so hosted endpoints (OpenAI, Azure OpenAI, GitHub Models, OpenRouter, Foundry) can't be used.
  • Only Ollama and OpenAI clients; no Azure OpenAI support (api-key header, api-version, deployment-based URLs).
  • IsEnabledCoreAsync sends a real "Are you there?" completion on startup, which costs money on paid endpoints.
  • languageModel schema describes the feature as "local language model".

Proposed changes

  1. Add apiKey to languageModel config, supporting @env:VAR_NAME so secrets stay out of config files; send as Bearer token (OpenAI-compatible) or api-key (Azure).
  2. Add AzureOpenAI client (or evaluate moving to Microsoft.Extensions.AI IChatClient).
  3. Replace the startup availability probe with a non-billable check (e.g. GET /models) where possible.
  4. Update rc.schema.json (current version folder) and docs.
  5. Never log API keys.

Follow-ups (out of scope here): per-plugin model overrides, redaction/opt-in for sending captured traffic to cloud models.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions