Routes to
Open models across many providers
Paid in
USDC on Solana
Receipts
Signed per call, rooted hourly
Router markup
None

The open inference router

Every open model.
One key.

One OpenAI-compatible API for the open models, one prepaid balance in USDC, and a signed receipt for every answer.

Routes to
  • llama-3.3-70b
  • qwen3-32b
  • deepseek-r1
  • mistral-small
  • gemma-3-27b
  • kimi-k2
  • glm-4.6
  • gpt-oss-120b
All models

An inference router you can check.

Trellis sits between your code and the providers that serve open models. It picks a provider, meters the call, charges what the provider bills and signs a receipt that says what happened.

Every hour the receipts are folded into one Merkle root and written to Solana, so a call can be checked later without taking our word for it.

Router markup on the price of inference
0%
Between receipt roots written to Solana
1 h
Signature on every receipt
Ed25519
Ways to pay: credits, per call or a tab
3
A garden trellis with climbing leaves, looking out over a valley at dawn

What every call gets.

The same three guarantees, whichever model or provider answers.

Compatible

Your SDK, unchanged

- baseURL: "https://api.openai.com/v1",
+ baseURL: `${TRELLIS_URL}/api/v1`,

Trellis speaks the OpenAI and OpenRouter APIs. Change the base URL and the key, and every model in the catalogue is one string away.

Browse the models

Accountable

A signed receipt

{
  "v": 2,
  "model": { "id": "llama-3.3-70b" },
  "node": { "provider": "<provider id>" },
  "lane": "public",
  "credit": {
    "mode": "prepaid",
    "cost_units": 812
  }
}

Each answer carries a receipt with the model, the provider, token buckets and the cost in millionths of a dollar, signed with Ed25519.

Check a receipt

At cost

One USDC balance

Prepaid credits
from 1 USDC
Per call, x402
0.03 USDC min.
Tab
up to 500 USDC

Prepay from any Solana wallet and spend on any model, or pay per call with x402. What is left goes back to the wallet that paid it.

Open the dashboard

Anchoring

Every hour, one root on Solana.

The hour’s receipts become a Merkle tree and its root is posted on chain. A short proof shows that any one receipt was in it.

How it works

The first call takes a minute.

Point the client you already use at Trellis. Read the quickstart

POST /api/v1/chat/completions
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: `${process.env.TRELLIS_URL}/api/v1`,
  apiKey: process.env.TRELLIS_KEY,
});

const reply = await client.chat.completions.create({
  model: "meta-llama/llama-3.3-70b-instruct",
  messages: [{ role: "user", content: "Hello" }],
});

Returned with the answer

x-receipt-id
The signed receipt for this call. Look it up on /verify/ or through the API.
x-generation-id
The generation record: model, provider, exact tokens and cost.
x-trellis-lane
The lane that served the call.
  1. Create a key

    Sign in with a Solana wallet or create a key. No email, no password.

  2. Add funds

    Top up from 1 USDC, or skip the balance and pay per call with x402.

  3. Call any model

    Use the SDK you already have. Each answer returns with its receipt id.

A garden gate set in a trellis wall, wrapped in climbing vines

Know where every request goes.

A request is served by one provider for the model you asked for, and the receipt names that provider. On the open lane the router reads the request in memory to route and meter it; the provider that answers reads it too.

Sealed lane

Confidential inference you can verify.

The sealed lane sends requests to a confidential gateway running on attested hardware, paid per token from the same balance.

Open the trust center
  1. A fresh attestation. The router reads the gateway’s attestation and verifies it.
  2. Pinned measurements. The measured software must match the values the router pinned; a change fails the check.
  3. A receipt that says so. The receipt carries the provider’s attestation hash beside the router’s signature.

Priced at cost.

Trellis adds nothing to the price of inference. Before a call, a hold covers the most it can cost; afterwards you are charged actual usage, never more than the hold.

Estimate a call
Router markup
None. You pay what the provider charges.
Prepaid credits
Top up from 1 USDC, from any Solana wallet.
Per call (x402)
0.03 USDC minimum. The Solana network fee is passed on at cost, on its own line.
Tabs
Up to 500 USDC, drawn per call and settled when the tab closes.
Launch limit
Up to 5,000 USDC of credits per key.
Leaving
Close the account and what is left goes back to the wallet that paid it.

What is live at launch.

Each feature outside the core joins only after its own security review. Take the full tour

Available at launch

  • Chat, completions and embeddings
  • The model catalogue and API keys
  • Prepaid credits, x402 per call and tabs
  • Signed receipts with hourly roots on Solana
  • The sealed lane, through an attested gateway

Enabled after review

  • The blind lane, paid with single-use tokens
  • Agent rulebooks and approvals
  • Teams, webhooks and batches
  • Bonded hosts on the open network

Questions.

The rest is in the documentation.

Read the docs
What is Trellis?

An inference router for open models, paid in USDC on Solana. It speaks the OpenAI and OpenRouter APIs, so the SDK you already use works, and every answer comes with a signed receipt.

Do I need an account?

No email and no password. Create a key, or sign in with a Solana wallet by signing a message. Signing in moves no funds; a top-up is its own transaction that you approve.

How is a call priced?

Before a call, a hold covers the most it can cost. Afterwards you are charged actual usage, never more than the hold. A provider failure that produced no tokens costs nothing.

What does a receipt prove?

That the router signed this record of the call: the model, the provider, the token buckets and the cost. The hourly Merkle root on Solana lets anyone show that the receipt is in that hour’s tree.

What happens if a provider fails?

The call can fall back to another provider serving the same model, and the receipt names the one that answered.

Who can read my prompts?

On the open lane the router reads the request in memory to route and meter it, and the provider that answers reads it too. The sealed lane sends it to a confidential gateway whose attestation the router checks first.

Can I withdraw my balance?

Yes. Close the account and what is left goes back to the wallet that paid it, at most what that wallet put in. The key stops spending as soon as you ask.

Start with one key.

Add a few dollars of USDC and call any open model.

A garden arch covered in vines, opening onto green hills at dawn