Skip to content

OpenAI Compatible

/Custom Provider

OpenAI Compatible

Any provider that implements /v1/chat/completions works out of the box. Groq, Together, OpenRouter, Mistral, Fireworks, Anyscale.

How it works

Rixel’s LLM client is OpenAI SDK compatible. If your provider returns choices[0].message.content and supports streaming SSE, it works. Just change baseURL and model.

Settings → Models → Add custom provider Base URL: https://api.groq.com/openai/v1 API Key: gsk_... Model: llama-3.3-70b-versatile
All these providers use same shape: bearer key + base URL ending in /v1. Rixel auto-detects streaming and falls back to non-streaming if needed.

Provider directory

Groq

Get API key →
baseURL https://api.groq.com/openai/v1
llama-3.3-70b-versatilemixtral-8x7b-32768llama-3.1-8b-instant

Fastest inference — ~400 tok/s. Great for low-cost scans. No tool calling on some models.

Together AI

Get API key →
baseURL https://api.together.xyz/v1
meta-llama/Meta-Llama-3.1-405B-Instruct-Turbomistralai/Mixtral-8x22B-Instruct-v0.1

Open models at scale. Use 405B for heavy reasoning.

OpenRouter

Get API key →
baseURL https://openrouter.ai/api/v1
openai/gpt-4oanthropic/claude-3-5-sonnetgoogle/gemini-pro-1.5meta-llama/llama-3.3-70b-instruct

Aggregator — one key, many providers, automatic fallbacks and routing.

Mistral

Get API key →
baseURL https://api.mistral.ai/v1
mistral-large-latestmistral-small-latestcodestral-latestopen-codestral-mamba

Codestral is excellent for fix PRs. La Plateforme API is OpenAI-compatible.

Anyscale

Get API key →
baseURL https://api.endpoints.anyscale.com/v1
meta-llama/Meta-Llama-3-70B-Instructmistralai/Mixtral-8x7B-Instruct-v0.1

Hosted open models, Ray-based autoscaling.

Fireworks AI

Get API key →
baseURL https://api.fireworks.ai/inference/v1
accounts/fireworks/models/llama-v3p3-70b-instructaccounts/fireworks/models/mixtral-8x22b-instruct

Very fast MoE models, function calling support.

OpenRouter — one key for everything

OpenRouter is most powerful for BYOK: single billing, automatic fallbacks, and you can mix models per severity.

// Rixel example: use Sonnet for critical/high, Mini for low
// Add two custom providers in Settings → Models:
1. openrouter-sonnet
   baseURL: https://openrouter.ai/api/v1
   model: anthropic/claude-3-5-sonnet

2. openrouter-mini
   baseURL: https://openrouter.ai/api/v1
   model: openai/gpt-4o-mini

// Future: per-severity routing in project settings (coming soon)
// For now, pick one as Preferred — switch before scan if needed.

// OpenRouter special: fallbacks via extra body
{
  "models": ["anthropic/claude-3-5-sonnet", "openai/gpt-4o"],
  "route": "fallback"
}

Groq — fastest scans

Groq LPU runs Llama 3.3 70B at ~400 tokens/sec. Perfect for triage phase — scan 100 files in seconds, then escalate critical findings to GPT-4o/Sonnet for fix generation.

curl https://api.groq.com/openai/v1/chat/completions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3.3-70b-versatile",
    "messages": [{"role":"user","content":"Find SQL injection in this code"}],
    "temperature": 0.1
  }'

Cost tip:

Groq has very generous free tier + $0.59 / 1M tokens for 70B. Use for initial pass, then use OpenAI/Claude only for findings that need fix PRs — saves 80% cost.
Supported models →Overview