Coding Assistant Config Generator

About

This tool generates ready-to-use configuration files and start commands for AI coding assistants. Works with any OpenAI-compatible endpoint — go-llm-proxy, vLLM, llama-server, Ollama, LiteLLM, or cloud APIs. Prefer not to edit config files? Use the "Start command" output to get a shell script you can run directly.

Everything runs client-side. No data is sent to, retained, or collected by any server.

Endpoint Setup

Models

Generate Configuration

Web Search

Local LLM backends don't support the web search tools that coding assistants request. go-llm-proxy solves this by intercepting search tool calls, executing them via Tavily or Brave Search, and injecting results back into the conversation — transparently, with no client-side setup.

With go-llm-proxy configured:

• Claude Code and Codex get native web search automatically — the proxy intercepts web_search tool calls and returns results in the format each client expects
• OpenCode and Qwen Code connect to the proxy's MCP endpoint for search, alongside any other MCP-compatible agent

Without a proxy, you can configure client-side search directly:

• Claude Code — add a Tavily MCP server via claude mcp add
• Codex — add a Tavily MCP server in config.toml
• OpenCode — add a Tavily MCP server in opencode.json
• Qwen Code — supports Tavily, Google Custom Search, and DashScope natively:

"webSearch": {
  "provider": [
    { "type": "tavily", "apiKey": "tvly-..." },
    { "type": "google", "apiKey": "...", "searchEngineId": "..." },
    { "type": "dashscope" }
  ],
  "default": "tavily"
}

DashScope is available automatically for Qwen OAuth users. Google requires a Custom Search API key and engine ID.

Images, PDFs & Vision

If you're pointing a coding assistant directly at vLLM or llama-server, image and PDF features silently fail on text-only models. You're limited to models with native vision support, or you lose screenshots, paste-image, and document reading entirely.

go-llm-proxy is a free, open-source single binary you run alongside your backend. It sits between your coding assistant and your inference server and handles this transparently:

• Images are routed to a vision model on your own hardware, described as text, and passed to your main model
• PDFs are text-extracted locally. Scanned pages fall back to an OCR model you choose
• User photos and tool output (screenshots, view_image) are handled separately — document-tuned models like PaddleOCR-VL process pages ~17x faster than general vision models
• Per-model config — use different vision/OCR processors per model, or disable for models that handle images natively

Everything runs on your machines. The proxy is a ~15MB binary with no dependencies, no cloud services, no accounts.