Open-source models, served on Bittensor

One API key for
open-source models.

An OpenAI-compatible endpoint over open-weight models. Change one line, keep your SDK. Fund your balance in USDG — no card, no invoice, no sales call.

See how it works

curl https://api.leafra.st/v1/chat/completions \
  -H "Authorization: Bearer $LEAFRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V3.2-TEE",
    "messages": [{"role": "user", "content": "Explain MEV in one sentence."}],
    "stream": true
  }'

Drops into anything that speaks OpenAI

OpenAI SDKVercel AI SDKLangChainLlamaIndexCursorContinue

Running in about two minutes

The slowest part is deciding which model to use.

01
Connect a wallet
Or sign in with email and we'll create a wallet for you. No forms, no company details, no approval queue.
02
Fund in USDG
Buy with a card, bridge USDC from any of 13 chains, or send USDG directly. Your balance is credited on confirmation.
03
Point your SDK at us
Swap the base URL, keep everything else. Streaming, tool calls and JSON mode work exactly as they did.

Boring where it counts

A gateway should be invisible. These are the parts that make it so.

Drop-in compatible

Same request shape, same response shape, same error envelope. If it works against OpenAI, it works here.

Streams properly

Responses pass straight through, never buffered. Time-to-first-token is the supplier's, not ours plus a delay.

No account friction

A wallet is the account. No card on file, no monthly minimum, no procurement call before your first token.

Per-key spend caps

Mint a key per project or per agent, cap what each can spend, and revoke one without touching the rest.

Built to fail over

Routing sits behind a supplier interface with health checks and retries, so adding capacity never touches your integration.

Usage you can audit

Every request records its model, token counts, latency and cost. Reconcile the bill yourself, line by line.

Pay for tokens. That's the whole model.

Top up a balance and spend it down. Nothing recurring, and nothing to cancel if you stop.

Starting from

$0/ month

Per-token rates are listed per model and shown before you send a request. You are charged the listed rate — never a percentage on top of an invisible one.

Pay only for tokens you use
No subscription, seats or minimums
Unlimited API keys, each with its own cap
Full usage history and cost breakdown
Balance never expires

Questions

Which models can I use?

Open-weight models served on Bittensor — the DeepSeek, Qwen, Llama and Mistral families among others. Call GET /v1/models with your key for the current list; availability tracks what the network is actually serving.

Do I need to understand crypto to use this?

No. Sign in with an email and you get a wallet without ever seeing a seed phrase, then top up with a card. The crypto is a payment rail, not something you have to think about.

Why USDG and not a card?

Settlement is instant, there are no chargebacks to price in, and it works the same in every country. That is why there is no monthly minimum — we are not amortising payment processing.

Is my prompt data stored?

We record metadata for billing — model, token counts, latency, cost. Prompt and completion bodies stream through and are not retained.

What happens if a model is down?

Routing retries against healthy capacity before returning to you. Errors surface in the OpenAI error shape, and responses carry headers showing which supplier served the request and how many attempts it took.

Can I use this from an autonomous agent?

Yes. Mint a dedicated key, set a spend cap on it, and revoke it independently of your other keys if the agent misbehaves.

One API key for open-source models.