Inference
that clears.

01 / Quickstart

First request in one line.

No SDK is required. The API accepts standard OpenAI Chat Completions JSON. Both hosts serve the same paths. Every completion requires an API key; the public model list does not.

curl https://api.verxl.com/v1/chat/completions \
  -H "Authorization: Bearer $VERXL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.5-4b-uncensored-64k",
    "messages": [{"role":"user","content":"Hello."}],
    "max_tokens": 128
  }'

The response is an OpenAI-shaped object. usage reports tokens; the provider is paid from the compute charge, not from the user's deposit.

{
  "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}
}

02 / Authentication

Public by default. Pay only for frontier.

Documentation and model discovery are public. Inference is never anonymous: every model, including Qwen 3.5 4B, requires a funded key.

Send a key only when you use a paid model:

Authorization: Bearer vrxl_live_...

Create and fund keys at verxl.com/console. An invalid key never silently falls back to an unauthenticated route; it returns 401.

03 / Models

Three routes. Published prices.

GET /v1/models is public and returns this list with prices and documentation links. Prices are per 1M tokens before the 50 bps protocol fee.

For OpenAI-compatible clients, the same context limit is returned as context_window, context_length and max_model_len, so generic model-list discovery does not fall back to 128K. DeepSeek off-peak is 50% off outside 01:00–04:00 and 06:00–10:00 UTC Monday–Friday, and all weekend. The 50 bps fee is charged on the compute amount and shown separately from the model price.

Qwen output floor. Qwen 3.5 4B uses an internal thinking pass. Requests with max_tokens below 256 are raised to 256 before dispatch so callers do not receive an empty content string with finish_reason: "length". The effective value is returned in verxl_clearing.effective_max_tokens and X-Verxl-Effective-Max-Tokens.

04 / Chat completions

One request shape. OpenAI compatible.

POST /v1/chat/completions accepts model, messages, temperature, max_tokens and stream. messages is required and must not be empty. Send Authorization: Bearer vrxl_live_... on every call.

Python

import os
from openai import OpenAI
c = OpenAI(base_url="https://api.verxl.com/v1", api_key=os.environ["VERXL_API_KEY"])
r = c.chat.completions.create(
    model="qwen3.5-4b-uncensored-64k",
    messages=[{"role": "user", "content": "Hello."}],
    max_tokens=128,
)
print(r.choices[0].message.content)

Create a key at verxl.com/console. The key is required even for Qwen 3.5 4B; there is no unauthenticated generation tier.

05 / Streaming

Server-sent events, end to end.

Set stream: true and read OpenAI-style data: lines. The stream ends with data: [DONE].

data: {"choices":[{"delta":{"content":"The fee is "}}]}

data: {"choices":[{"delta":{"content":"50 bps."}}]}

data: [DONE]

06 / Errors

Failures are structured.

Every non-2xx response uses the same envelope:

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

07 / Settlement

50 bps, not a subscription.

Verxl is non-custodial. The protocol fee is 50 bps of the compute charge (0.5%). The provider receives 99.5% in USDC on Arbitrum. There is no monthly platform fee and no minimum deposit.

Sepolia is a test deployment. Mainnet production is out of scope for this build.

08 / MCP

Agent-native setup.

Verxl exposes a remote MCP server at https://verxl.com/mcp. It is designed for an AI agent to discover the API, explain the paid setup to its user, and then call inference once the user supplies a funded key.

The MCP server never holds funds, private keys or API keys. It cannot sign a deposit or mint a funded key; it gets the agent to the exact console step and then the user approves. An inference call without a key returns 401 with a machine-readable help object pointing back to the console, docs and MCP.

https://verxl.com/mcp
transport: streamable-http
auth: none for setup/discovery; vrxl_live_... for verxl_chat

09 / For agents

Read the spec, not the SPA.

Agent-facing resources are static, public and do not need a browser:

The API returns CORS * and supports browser clients. Public reads need no key; inference always does. Do not send a placeholder key unless your SDK requires a non-empty value.