# Verxl Protocol — Inference API

> OpenAI-compatible inference routing and non-custodial clearing. The protocol takes 50 bps; the provider keeps 99.5%, settled in USDC on Arbitrum. Documentation, model discovery and health are public. Every inference request requires `Authorization: Bearer vrxl_live_...` and a funded balance; there is no free compute tier.

## Start here (no authentication)

- Docs: https://verxl.com/docs
- OpenAPI 3.1: https://verxl.com/openapi.json
- Models: https://verxl.com/v1/models
- Health: https://verxl.com/api/health
- API host: https://api.verxl.com (same paths)
- Apex host: https://verxl.com

## Public discovery (no key)

These resources are public and safe for an AI agent to inspect:

- `GET https://verxl.com/v1/models` — model ids, context, published prices, docs links
- `GET https://verxl.com/openapi.json` — OpenAPI 3.1
- `GET https://verxl.com/llms.txt` — this file
- `GET https://verxl.com/api/health` — edge + protocol health

No key is required for these reads. No completion request is free.

## Authentication for inference

```
Authorization: Bearer vrxl_live_...
```

Every completion requires a key and a funded balance. Create and fund one at https://verxl.com/console. A missing or invalid key returns `401`. Models:
- `deepseek-v4.1-flash` — 1M context, official bridge
- `mimo-v2.6-flash` — 1M context, official bridge

## Models

| id | context | prompt $/1M | completion $/1M | access |
|---|---:|---:|---:|---|
| qwen3.5-4b-uncensored-64k | 65,536 | 0.05 | 0.10 | key + balance |
| deepseek-v4.1-flash | 1,000,000 | 0.30 | 1.20 | key + balance |
| mimo-v2.6-flash | 1,000,000 | 0.14 | 0.28 | key + balance |

DeepSeek off-peak pricing is 50% off outside 01:00–04:00 and 06:00–10:00 UTC Monday–Friday, and all weekend. The 50 bps protocol fee is separate and is deducted from the compute charge.

## Request shape

```json
{
  "model": "qwen3.5-4b-uncensored-64k",
  "messages": [
    {"role": "system", "content": "Answer directly."},
    {"role": "user", "content": "Hello."}
  ],
  "temperature": 0.7,
  "max_tokens": 128,
  "stream": false
}
```

`messages`, `model`, `temperature`, `max_tokens` and `stream` follow the OpenAI Chat Completions API. Use an exact model id from `/v1/models`; unknown ids fall back to the default route.

## Response shape

```json
{
  "id": "chatcmpl_01JVERXL",
  "object": "chat.completion",
  "created": 1762000000,
  "model": "qwen3.5-4b-uncensored-64k",
  "choices": [{
    "index": 0,
    "message": {"role": "assistant", "content": "..."},
    "finish_reason": "stop"
  }],
  "usage": {"prompt_tokens": 24, "completion_tokens": 23, "total_tokens": 47}
}
```

SSE responses use OpenAI `data:` chunks and terminate with `data: [DONE]`.

## Errors

All non-2xx responses use:

```json
{"error":{"message":"...","type":"invalid_request_error","code":"bad_json"}}
```

Common statuses: `400` invalid request, `401` missing/invalid Authorization, `402` insufficient balance, `429` rate limit, `500` upstream failure.

## MCP

Remote MCP server: `https://verxl.com/mcp` (Streamable HTTP, JSON responses).

Tools:
- `verxl_get_started` — no key; exact paid setup and URLs
- `verxl_list_models` — no key
- `verxl_get_api_key` — no key; returns console URL and user steps; cannot create/fund a key
- `verxl_payment_instructions` — no key; treasury, chain, token, minimum; cannot sign
- `verxl_chat` — requires a funded `vrxl_live_...` key passed as an argument

The server never holds funds, keys or private keys. The user approves the wallet deposit in the console. There is no free compute tier.

## CORS and agents

- `Access-Control-Allow-Origin: *`
- `GET`/`OPTIONS` on `/v1/models`
- `POST`/`OPTIONS` on `/v1/chat/completions`
- Every inference request needs a key; the public reads above do not.
- For agent use, read `/openapi.json` first and call `/v1/models`; do not scrape the SPA.
- Settlement is non-custodial; Verxl does not hold user funds.

## Human links

- Product: https://verxl.com
- Chat: https://verxl.com/chat
- Console: https://verxl.com/console
- Source: https://github.com/simonhavlik/verxl-protocol