Get started

Quickstart

Source

Every read endpoint on vitna.ai is public and needs no key. There is no signup, no token, and no rate limit to negotiate. If you can reach the host you can read the record.

The first request

bash
curl https://vitna.ai/api/v1/markets

That answers the component classes Vitna prices, how many skus and providers each one carries, and how fresh the freshest and stalest figures in it are.

The tape for one class

bash
curl https://vitna.ai/api/v1/markets/model/tape

One row per priced cell: a sku, a dimension, an amount in USD, the unit that amount is quoted in, when the price started, when it was last confirmed, and the sha256 of the archived document it was read from.

Valid classes are model, tooling, retrieval, extraction, translation, speech and compute. Every declared class is wired as of 2026-09-10. A class that is registered and not yet read answers an empty tape rather than a 404, and the markets endpoint keeps it in the answer at zero.

One sku over time

bash
curl "https://vitna.ai/api/v1/skus/gpt-5-pro/history?dimension=input_per_mtok"

Price periods, oldest first. A period is a stretch during which the price did not move.

The spread on one product

bash
curl https://vitna.ai/api/v1/spreads/compute/h100-sxm

The same product sold by several providers, cheapest first, with the percentage between the low and the high.

Subscribe to the log

bash
curl https://vitna.ai/api/v1/feed.json

JSON Feed 1.1. Price moves, launches, policy changes and deprecations, plus the triggers that fired and the decisions the engine proposed.

Check a figure yourself

Take an observation id from the record and ask for its proof:

bash
curl https://vitna.ai/api/v1/observations/1/proof

You get the leaf's own fields, its hash, the sibling hashes up the Merkle tree, and the signed root the walk should arrive at. Rebuild the hash, fold the path, verify the Ed25519 signature. Nothing in the answer has to be trusted for the check to mean something: if any of it were wrong, the arithmetic would stop agreeing.

The full walkthrough, both checks end to end, is in Verify a figure yourself.

Three-minute gateway quickstart (LiteLLM and OpenRouter)

Teams already running LiteLLM or OpenRouter in production can adopt Vitna's attested market prices, local unit profiles, and signed routing decisions in three steps without swapping out their gateway first.

Step 1: Load attested Witness prices in LiteLLM (no key, no signup)

bash
curl -s https://vitna.ai/api/v1/gateways/litellm/model-prices.json | head -c 400
bash
LITELLM_MODEL_COST_MAP_URL=https://vitna.ai/api/v1/gateways/litellm/model-prices.json \
  litellm --config litellm.yaml

Every entry translates Witness input_per_mtok, output_per_mtok, and embed_per_mtok figures into per-token USD alongside a vitna provenance block (sku, sourceUrl, confirmedAt, evidenceSha256, citationUrl, and full context bands). Responses carry ETag, so polling with If-None-Match returns 304 when nothing changed. See LiteLLM cost map.

Step 2: Measure your own token unit profile locally

Run sample or mirrored traffic through Relay (no policy, no vitna.ai account, nothing leaves your machine), then compute your observed token profile from its append-only JSONL ledger:

bash
OPENAI_API_KEY=sk-... npx -p vitna-runtime vitna-relay
vitna-runtime-profile --file ~/.vitna/relay/ledger.jsonl
vitna-runtime-profile --file ~/.vitna/relay/ledger.jsonl --json

vitna-runtime-profile prints --profile inputTokensPerUnit=... --profile outputTokensPerUnit=... for each lane, separates refusals and unrouted passthrough rows rather than estimating missing counts, and stays completely offline. Operators recording into an Aperture register can also extract p90 latency and sample counts directly via npm run aperture:evidence:from-run (Measuring your unit profile, Candidates and evidence).

Step 3: Feed CI evals and render signed policies back to your gateway

Send existing Promptfoo, Braintrust, or UK AISI Inspect AI results directly to the append-only evidence endpoint (CI evaluation evidence) or enter them via npm run aperture:evidence:add (Candidates and evidence):

bash
curl -X POST https://vitna.ai/api/v1/org/evidence \
  -H "authorization: Bearer vitna_key_..." \
  -H "content-type: application/json" \
  -d '{
    "adapter": "promptfoo",
    "workload": "deep-research",
    "sku": "openai/gpt-5.6-luna",
    "label": "pr-412-gpt-5.6-luna",
    "sourceRef": "https://github.com/convexity/app/actions/runs/91823",
    "report": { "results": { "stats": { "successes": 193, "failures": 7, "errors": 0 } } }
  }'

Once a workload decision is accepted and signed into policy.json, render that same policy into native configuration for your existing proxy (litellm, openrouter, cloudflare, or portkey) without changing your serving data plane (Gateway adapters):

bash
VITNA_PUBLIC_KEY=$(cat policy-key.pub) \
  npm run aperture:adapter -- --target litellm --policy policy.json --out litellm.yaml
VITNA_PUBLIC_KEY=$(cat policy-key.pub) \
  npm run aperture:adapter -- --target openrouter --policy policy.json --out presets.json

Conventions worth knowing before you parse anything

  • Every response is a JSON object with a meta key. See api/v1/overview.md.
  • meta.source tells you whether the answer came from the live record (database) or from the snapshot committed in the repository (snapshot). The public site answers from the record.
  • Money is USD, quoted in the unit the dimension declares, stored to six decimal places.
  • Timestamps are RFC 3339. Calendar dates are YYYY-MM-DD and mean the day.
  • Responses carry an ETag. Send it back as If-None-Match and you will get a 304 when the record has not changed.

What this API will never do

It does not route traffic, hold your keys, proxy inference, or take an action on your behalf. Vitna reads published prices and records decisions about them. Changing what your stack runs is a person's job in another system.