gateway online

OpenAI-compatible
est. 2026

nara@api ~ %

Point your SDKat every model.

An OpenAI-compatible gateway for chat and image generation. One Bearer key, one credit meter, 20+ models — change the base URL, keep your code.

Get your key — 100 credits freeRead the docs

nara — curl

$ curl api.naravirtual.ai/v1/chat/completions \ -H "Authorization: Bearer $NARA_KEY" \ -d '{ "model": "claude-3-5-sonnet", "stream": true }'

data: {"delta":{"content":"One key,"}}data: {"delta":{"content":" every model."}}data: {"usage":{"prompt_tokens":21,"completion_tokens":128}}data: [DONE]

── meter · reserved 12 → committed 8 → refunded 4 · balance 9,992

Scroll

/v1/chat/completions · /v1/image/generations · credits · webhooks

GPTClaudeGeminiGrokLlamaMistral

/v1/chat/completions/v1/image/generationscredits ledgersigned webhooksBearer authSSE streaming

/manifesto

positioning · v2

/lifecycle

Every request, four moves.

01reserve

A hold lands before a single token streams.

The gateway reserves credits at the model's catalog ceiling the instant your request arrives. Concurrent calls can't overdraw the balance — the meter is race-safe by construction, not by retry.

02route

One key fans out to every provider.

The catalog resolves your model string to a provider, checks kill-switches, and arms the fallback chain. BYOK keys pass straight through at zero markup — your credits stop counting entirely.

03generate

Deltas stream back OpenAI-shaped.

Server-sent events arrive in the exact shape your SDK already parses — content deltas, tool calls, usage. Token counts accumulate live against the hold, never past it.

04commit

The ledger keeps only what you used.

On completion the actual token cost commits, the rest of the hold releases back to your balance, and an idempotent ledger row is written. Retries can't double-charge. Ever.

nara — reserve

nara chat --reserve --model claude-3.5-sonnet

→ safety hold: 12 credits

→ balance verified · concurrent-safe

balance 10,000.00 cr

hold − 12.00 cr ← catalog ceiling

available 9,988.00 cr

nara — route

route --provider auto

→ one key, every provider

→ fallback chain armed

anthropic ● selected byok: off

openai ○ fallback-1 armed

google ○ fallback-2 kill-switch: clear

nara — generate

stream

→ 842 tokens · tool calls rendered

→ memory injected · web search off

tokens 842 cost-so-far 7.9 cr

nara — commit

commit --actual 8

→ 8 credits committed

→ 4 refunded · usage logged

hold 12.00 cr

commit − 8.00 cr

release + 4.00 cr refunded

balance 9,992.00 cr ✓ idempotent

01reserve

A hold lands before a single token streams.

The gateway reserves credits at the model's catalog ceiling the instant your request arrives. Concurrent calls can't overdraw the balance — the meter is race-safe by construction, not by retry.

nara — reserve

nara chat --reserve --model claude-3.5-sonnet

→ safety hold: 12 credits

→ balance verified · concurrent-safe

balance 10,000.00 cr

hold − 12.00 cr ← catalog ceiling

available 9,988.00 cr

02route

One key fans out to every provider.

The catalog resolves your model string to a provider, checks kill-switches, and arms the fallback chain. BYOK keys pass straight through at zero markup — your credits stop counting entirely.

nara — route

route --provider auto

→ one key, every provider

→ fallback chain armed

anthropic ● selected byok: off

openai ○ fallback-1 armed

google ○ fallback-2 kill-switch: clear

03generate

Deltas stream back OpenAI-shaped.

Server-sent events arrive in the exact shape your SDK already parses — content deltas, tool calls, usage. Token counts accumulate live against the hold, never past it.

nara — generate

stream

→ 842 tokens · tool calls rendered

→ memory injected · web search off

tokens 842 cost-so-far 7.9 cr

04commit

The ledger keeps only what you used.

On completion the actual token cost commits, the rest of the hold releases back to your balance, and an idempotent ledger row is written. Retries can't double-charge. Ever.

nara — commit

commit --actual 8

→ 8 credits committed

→ 4 refunded · usage logged

hold 12.00 cr

commit − 8.00 cr

release + 4.00 cr refunded

balance 9,992.00 cr ✓ idempotent

/endpoints

The surface you call

full reference

POST/v1/chat/completions

Chat completions

OpenAI-compatible chat with streaming and function calling. Swap the base URL, keep every line of your SDK code — one model param reaches 20+ models.

{ "model": "claude-3-5-sonnet",

"messages": [...], "stream": true }

streamingfunction calling20+ models

POST/v1/image/generations

Image generations

Prompt to pixels with per-op credit costs. Sizes up to 1024×1024, batched n, same Bearer key.

{ "prompt": "…",

"size": "1024x1024" }

1024×1024per-op cost

GET/credits

Credits & keys

Balance, per-key spending caps, and an audit ledger for every reserve, commit, and release. Mint unlimited keys — name, scope, and cap each one.

{ "balance": 9992,

"reserved": 0, "keys": 4 }

per-key capsaudit ledger

EVENT/webhooks

Signed webhooks

HMAC-SHA256 signatures on every delivery, unique delivery IDs for de-duping, and up to 5 attempts with backoff when your endpoint is down.

X-Nara-Signature: sha256=…

X-Nara-Delivery-Id: dlv_…

X-Nara-Signature5 retries

Models routed

AI products

Credit meter

/trust

Built so you can leave. You won't.

01base_url swap

OpenAI-compatible

The official OpenAI SDKs work by changing base_url — nothing else. No proprietary client, no lock-in, leave any time with zero rewrites.

020% markup

BYOK passthrough

Bring your own provider keys and requests route through them at cost. Credits stop counting; your keys live in an encrypted vault and never leave the gateway.

03race-safe meter

Idempotent billing

Reserve → commit → release. Failed calls release the full hold, retries can't double-charge, and every credit movement lands in an auditable ledger.

/start

First token in under a minute.

100 free credits, no card. One curl away from every model.

Get your keyRead the docs