Provider

A provider is a concrete service that implements a capability.

If image.generate is the capability (“the framework can turn a prompt into an image”), then providers are who actually does it:

• a local ComfyUI workflow on your machine,
• an Anthropic API call,
• an OpenAI DALL·E call,
• a self-hosted Stable Diffusion server,
• a community-built adapter pointing at the next thing.

Your project code never names a provider directly — it asks the capability. Configuration decides which provider runs.

How a provider gets registered

A provider declares:

• Which capability it implements. A single provider can implement more than one (an LLM-with-vision provider implements both llm.call and image.understand).
• Its inputs and outputs. Provider adapters normalise to the capability’s wire shape so consumers don’t see provider-specific fields.
• Cost + throughput. Per-call cost, rate limits, latency estimates.
• Auth. API key location, OAuth flow, BYOK header, etc.

The four-tier priority order

When multiple providers can implement the same capability, the framework picks in this order:

1. Local — runs on your machine. ComfyUI, llama.cpp, a local embedding model. Free, private, no rate limits.
2. Free — API tiers with a free quota (Groq, some Hugging Face inference endpoints).
3. Pay-per-use — per-call billing (Anthropic, OpenAI, Gemini direct).
4. Subscription — Claude Code Max, GitHub Copilot, OpenAI Codex CLI. The framework can route through these if the user already pays for them, avoiding double-billing.

The framework respects the order automatically. You can override it in ~/.config/vibesmith/config.toml if you have a reason.

Why this matters

You don’t write your project code to “use ComfyUI” or “use Anthropic.” You write it to use the capability. That means:

• Switching providers doesn’t touch your code. Change one config line; the framework re-routes.
• A teammate without the local provider can still run your project — the framework falls back to a paid provider with a warning.
• The AI assistant can suggest provider swaps. “You’re paying $X/month on Anthropic; here’s the local provider that would handle this capability.”

Provider adapters are pluggable

The framework ships a handful of native adapters (Anthropic, OpenAI, Gemini, ComfyUI). Community adapters extend the set via the same ProviderAdapter contract — no framework changes required to add a new one.

If you want to write an adapter, you implement the contract for one capability, declare your cost + throughput, and the framework slots you into the routing tier.

Next

• Capability — the abstract surface providers implement.
• First-party chat panel — the BYOK chat surface that uses LLM providers.
• AI assistant — the larger AI-integration story.