AI Provider System

MIFY provides unified access to 140+ AI providers / 2,600+ models / 35 unique endpoints via the LiteLLM gateway, plus 16 first-class native integrations. (LiteLLM adds providers continuously — see the canonical link above for the live count.)

How It Works

Your Workflow
  → MIFY selects provider based on pack/tier
    → AI SDK adapter normalizes the request
      → Provider receives the call
        → Response normalized back to MIFY format

You write your workflow once. MIFY handles provider differences — authentication, request format, response parsing, error handling, and cost tracking.

Provider Tiers

Provider packs group models by optimization target:

Tier	Optimized For	Example Models
Fast	Low latency	GPT-4o-mini, Claude Haiku, Gemini Flash
Balanced	Cost/quality	GPT-4o, Claude Sonnet, Gemini Pro
Accurate	Best quality	GPT-4, Claude Opus, Gemini Ultra

Select a tier when running workflows, or configure a default.

Local Models (Ollama)

Run AI models locally with zero API costs:

1. Install Ollama: curl -fsSL https://ollama.com/install.sh | sh
2. Pull a model: ollama pull phi3:mini
3. In MIFY, select Ollama as your provider — no API key needed

Cloudflare Workers AI

Run models at the edge with Cloudflare:

• Chat (LLaMA, Phi)
• Embeddings (BGE)
• Image Generation (Stable Diffusion XL)
• Vision, Speech Recognition, Text-to-Speech
• Translation, Classification, Object Detection

Cost Tracking

MIFY tracks AI costs per workflow run:

• Token usage per node
• Cost estimation based on provider pricing
• Usage dashboard at /settings/usage
• Admin usage overview at /admin/usage

BYOK (Bring Your Own Key)

Add your own API keys for any provider:

1. Go to Settings → Credentials
2. Select the provider
3. Enter your API key
4. The key is encrypted at rest and used for your workflows only

Cloudflare AI Gateway (optional)

Admins can route LiteLLM traffic through a Cloudflare AI Gateway for caching, rate-limiting, and analytics — toggled in /admin/gateway.

Cloudflare Workers AI Fallback

If the primary provider is down or over-quota, MIFY can fall back to Cloudflare Workers AI automatically — chat, embeddings, and image generation continue working without operator intervention.

• Three-stage fallback — flat settings (per-workspace toggle) + kill switch (org-wide off) + routing (per-route fallback model selection)
• Admin emergency disable — admin-only /api/admin/cf-killswitch/* API engages or clears the global kill switch (no dedicated UI page yet — invoked from /admin/gateway or via API)
• Configurable per workspace — each workspace can opt in/out and pick which CF Workers AI model to use as the fallback target
• Auto-wired — when a workspace has CF fallback configured (Option Y), ProviderResolver wires it into every LLM call
• Embedding support — Cloudflare embeddings work as a fallback for RAG templates whose primary embedding provider is unavailable

Per-Workspace Backends (Backend Capability Registry)

Beyond LLMs, every runtime backend has a per-workspace setting:

Capability	Setting Page	What It Picks
LLM	/workspaces/[id]/settings/llm	Which provider key/region/model to use
Document Parser	/workspaces/[id]/settings/parser	Local vs Unstructured.io sidecar for PDF/DOCX
Browser	/workspaces/[id]/settings/browser	Browser automation backend
Sandbox	/workspaces/[id]/settings/sandbox	RawHost vs Cloudflare Sandbox for code exec

Each capability also has an admin fallback page (/admin/{llm,parser,browser,sandbox}-fallback) for org/global defaults when no workspace setting exists.