sarmalink-ai

SarmaLink-AI

Star History

An open-source, multi-provider AI backend with automatic failover.

Built by Sarma Linux — 17 months of development, open-sourced for everyone.

🚀 v1.1.0 shipped on 2026-04-20 — OpenAI-compatible proxy at /api/v1/chat/completions, four new free-tier providers (GitHub Models · Cohere · Mistral · Ollama local fallback), per-step fallback metrics, 110 tests green. See CHANGELOG.md for the full list and docs/OPEN-ISSUES.md for what's open for contributors.

What this repo is: a headless Next.js backend — the failover engine, auto-router, live tools, Supabase schema, and REST/SSE API (/api/ai-chat). It is not a drop-in ChatGPT clone UI. If you want the full hosted product with dark mode, markdown rendering, and thinking traces, see sarmalinux.com/products/sarmalink-ai. To build your own UI on top, point any chat client at the SSE endpoint — see docs/HOW-IT-WORKS.md for the streaming protocol.

---

What is SarmaLink-AI?

SarmaLink-AI is a production-ready AI chat backend that routes every message through a failover of AI engines. If one engine is busy, the next fires in under 50 milliseconds. Users never see errors — they always get an answer.

• 36 AI engines across 7 providers (Groq, SambaNova, Cerebras, Google Gemini, OpenRouter, Cloudflare, Tavily)
• Smart auto-routing — detects whether you're asking for code, web search, quick answers, or deep reasoning, and picks the right model automatically
• Live tools — real-time exchange rates (ECB), weather (any city), container/shipment tracking (9 shipping lines)
• Image generation & editing — FLUX.2 klein (instruction-following, not noise-based)
• Persistent memory — learns who you are across conversations (like ChatGPT memory)
• Document analysis — upload PDFs, Excel, Word files (up to 10 at once)
• 50 saved conversations per user, oldest auto-deleted
• SSE streaming API with separate token and thinking event types — bring your own UI, or consume from any HTTP client

Total cost: the providers' free tiers. No credit card needed for any of them.

---

Architecture

User sends message
    ↓
Auto-router (regex, instant) → Smart / Reasoner / Live / Fast / Coder / Vision
    ↓
Live tools fire first (exchange rates, weather, tracking)
    ↓
Failover engine (up to 14 steps):
    Step 1: SambaNova DeepSeek V3.2 (685B) → try all keys
    Step 2: Groq GPT-OSS 120B → try all keys
    Step 3: Llama 4 Maverick → try all keys
    ...
    Step 14: OpenRouter Nemotron Nano 30B :free
    ↓
Each step: if 429/5xx → next step in <50ms
    ↓
Stream tokens via SSE → render markdown → save to Supabase
    ↓
Extract memories in background (Groq Llama 8B)

---

The 6 Modes