One account. One endpoint. Every model VG Code needs.
Vibgrate Relay is the hosted model router for VG Code: curated hosted models behind one OpenAI-compatible endpoint, with exact per-token metering, org-level governance, and BYOK passthrough. It's the default hosted option — OpenRouter and other providers remain fully supported alongside it, and local Code Modes remain the fully offline path.
Reserve, serve, settle
A request's cost is estimated and held against your balance before anything is sent upstream. Once the model reports real usage, Relay settles to the exact cost and returns the unused hold — you are never charged for tokens you didn't get.
$ export VIBGRATE_RELAY_TOKEN="<dsn-key>:<dsn-secret>" $ vg code "extract the pagination logic in UserList.tsx into a hook" Vibgrate Relay — VIBGRATE_RELAY_TOKEN is set → anthropic/claude-sonnet-5 Reserved $0.0421 (estimate) Served 289 files considered · graph-grounded diff proposed Settled $0.0113 (actual) · $0.0308 returned to balance 1 file changed, +34 -12 Review the diff? [y/N]
Illustrative example — actual reserve/serve/settle values depend on the model, prompt, and current rate card.
Built for teams shipping with AI
One endpoint, curated models
No provider-hopping. Vibgrate Relay serves a curated, priced catalog of hosted models through one OpenAI-compatible endpoint — served directly off Cloudflare Workers AI where possible, proxied to the provider otherwise, at the same quoted price either way.
You pick the model, once
Relay never auto-switches your model mid-session based on cost or latency. You choose it once, for the whole run — the same predictable model identity from the first token to the last.
Exact per-token metering
Prepaid credits, reserved before a request is served and settled to the real cost afterward — never a surprise invoice. A failed request is never charged. Published fee schedule, no hidden markup in a rounded number.
Org-level governance
Model allowlists, spend caps (workspace and per-request), and residency pinning — configured once for the whole team, not left to every seat to police individually.
BYOK passthrough
Bring your own provider key and route through Relay for governance and audit at no Relay-credit cost — free below a generous monthly floor, with the exact same context quality as the paid lane.
A router, never a model
Vibgrate doesn't train or host frontier models. Relay is the metering, governance, and billing layer in front of the models you choose — it never touches what VG Code proposes, only how the completion behind it is served and billed.
Prepaid, published, no surprises
Relay bills prepaid credit by default — nothing is served on Vibgrate's key unless your balance can already cover it, so there's no invoice shock. Top up from $25; a processing fee applies (from 3.9% on $100+ top-ups, 4.9% below), always shown as a visible line item, never hidden in a rounded number. Bring your own provider key instead, and pay nothing beyond a generous monthly floor.
Prepaid credits
$25–$1,000 top-ups, 12-month expiry, FIFO consumption by expiry, optional auto top-up with daily caps.
BYOK passthrough
Your own provider key, routed through Relay for governance — free below $30k/mo (pay-as-you-go) or $250k/mo (enterprise), 4% marginal fee above.
Get started
Install the Vibgrate CLI, sign in, and point VG Code at Relay.
npm install -g @vibgrate/cli vg login export VIBGRATE_RELAY_TOKEN="<your-dsn-key>:<your-dsn-secret>" vg code "your first instruction" --provider vibgrate-relay
No credit? Use your own provider key or --local Code Modes instead — both are free.
What Relay doesn’t do
We don’t train or host models
Relay is a routing, metering, and governance layer in front of models you choose — never a model itself.
We don’t auto-swap your model
No cost-based, latency-based, or usage-based routing mid-session. You pick the model once, for the whole run.
We don’t charge for failed requests
An upstream error releases the full reservation before returning — you pay for tokens delivered, never requests attempted.
We don’t downgrade BYOK
Bringing your own key gets the exact same context quality as the paid lane — no capability withheld to make Relay look better.
Route your first VG Code session
vg doctor shows exactly which backend and model VG Code will pick, and why — confirm your setup before the real run.
Keep reading
Vibgrate Relay is a routing service. It does not train, fine-tune, or claim authorship of any model's output — model responses are produced by the underlying provider or serving backend selected for your request.