The runtime
An optional process that applies an accepted decision per request, on your own infrastructure. It is the second way to consume a decision, beside the config a consumer loads into a variable it already reads.
The distinction the whole thing rests on
Vitna's software may sit in your request path. Vitna's service may not.
The runtime loads a signed policy from local disk at boot and answers every request from it. It makes no call to vitna.ai: not at boot, not per request, not on a timer. There is no endpoint that serves a policy to a process, and there will not be one.
So Vitna can be entirely unavailable and every request keeps routing, because the answer to "where should this request go" was signed into a file before the request existed.
What it is for
A team with a gateway already has somewhere to put a decision, and config mode costs them no runtime footprint at all. A team without one has nowhere to load an environment fragment into. The runtime is the same decisions, applied.
It also closes a gap the record has: a usage export read after the fact is scoped by whatever the exporting system could scope it by, which is rarely the lane. A ledger written at the moment of the request is lane-exact because the caller named the lane.
What it will not do
| It never | Because |
|---|---|
| Calls vitna.ai | A runtime that fetched its policy would put Vitna's availability inside yours. Pushing the ledger is a separate command an operator schedules, and the serving process holds no code path to vitna.ai with push enabled or without |
| Substitutes a model | A lane that cannot be reached is an error you see, not a quiet reroute |
| Estimates a token count | A provider that reported no usage is one whose usage the record does not know |
| Serves what it cannot attribute | Every answer has a lane, a decision and a policy digest behind it, or is marked unrouted. Relay marks every answer unrouted, because it has no policy and so no lane |
| Sends your prompts anywhere but the provider | Your keys, your accounts, your traffic |
| Retries, or serves past a chain | A second serve is a fallback a person registered against the lane's decision, walked in order and recorded per attempt; nothing is ever served that no chain names. This row said "never fails over" until R2 shipped chains, and went stale for a day |
The pieces
| Page | What it covers |
|---|---|
| Relay: the runtime with no policy | The same runtime with no policy, serving every request as the model it names, for a developer with no register yet |
| Per-request routing with vitna-route | Deterministic, cache-sticky per-request model and effort routing (shadow first, promoted by an Aperture-signed policy) |
| A lane on your own machine | One command from a workload in your register to a runtime on localhost |
| Installing the runtime | npm, the container, and Deploy on a cloud account you own |
| Serving a lane | Starting it, the request it takes, and what it refuses |
| Redaction | Lifting known credential formats out of a request before it leaves, and putting them back into the answer, where you turn it on |
| The response cache | Answering a repeat of a deterministic request from memory, where you turn it on, and writing it down as no provider's work |
| Putting a coding harness on a lane | One command that starts opencode or Crush against a lane, with no provider key in its environment |
| The ledger | What one request records, and reading it back into the register |
| Measuring your unit profile | What one call on a lane consumes, measured from that ledger, locally |
| Pushing the ledger | The opt-in command that sends closed days to the record |
