Overview
Vibgrate Relay is Vibgrate's hosted model router. It gives VG Code — and other Vibgrate AI surfaces — one account and one OpenAI-compatible endpoint for reaching curated hosted models, with exact per-token metering, org-level governance (model allowlists, spend caps, residency pinning), and BYOK (bring-your-own-key) passthrough.
Relay is a routing service, never a model. Vibgrate does not train or host frontier models. Relay serves curated third-party models (via Workers AI where possible, otherwise proxied to the provider) and — where Vibgrate has direct capacity — its own inference paths. Think of it as the metering, governance, and billing layer that sits in front of the models you already choose.
Why it exists
Running VG Code against a hosted model today means picking a provider yourself — OpenRouter, OpenAI, Anthropic, Together, LiteLLM — each with its own account, its own API key, its own billing dashboard. That's fine for an individual, but it falls apart the moment a team wants:
- One bill instead of N provider invoices to reconcile.
- One place to cap spend, instead of trusting every seat to police their own usage.
- One allowlist of which models are approved for use against company code, instead of a policy nobody can enforce.
- A hosted default that still respects BYOK — bring your own provider key when you have one, without losing governance and audit.
Relay is that layer. It sits between VG Code and the model, never between you and your code — VG Code is still dry-run by default and never writes without explicit consent, regardless of which backend serves the completion.
How it fits with everything else
| Surface | What it is | Relation to Relay |
|---|---|---|
VG Code (vg code) | The graph-grounded, governance-gated code-editing capability in the Vibgrate CLI | Relay is the default hosted backend once live; local Code Modes and other providers (OpenRouter, OpenAI, etc.) remain fully supported |
Vibgrate AI Context (vg serve) | The local-first MCP server — code map, drift, and library docs | Unrelated to model hosting; Relay never touches what vg serve returns, only how a hosted completion is served and billed |
| Vibgrate MCP | The hosted MCP server over your Vibgrate Cloud data | A different hosted surface — MCP serves your data, Relay serves model completions |
| Vibgrate Cloud | The team dashboard | Relay's billing, usage, and governance controls live under Vibgrate Cloud → Settings → Billing → Relay |
What Relay is not
- Not a model. Relay never trains, fine-tunes, or claims authorship of a model's output.
- Not a per-request router. Relay does not silently swap your model mid-session based on cost or latency — see Model selection & routing.
- Not mandatory. OpenRouter and other providers remain supported alongside Relay, and local Code Modes remain the fully offline path with no Vibgrate account required.
Next steps
- Get started with Vibgrate Relay — set the token, pick a model, run
vg code. - Supported models & catalog — what's servable today and how the catalog stays current.
- Pricing & credits — prepaid credits, top-up sizes, and the published fee schedule.