API (Anthropic-compatible)

Quickstart (cURL)

curl https://grid.x402compute.cc/v1/messages \
  -H "x-api-key: $SGL_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "qwen-2.5-32b",
    "max_tokens": 256,
    "messages": [{"role": "user", "content": "Explain confidential compute in one line."}]
  }'

With the Anthropic SDKs

# pip install anthropic
import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://grid.x402compute.cc",
    api_key=os.environ["SGL_API_KEY"],   # a Singularity Layer compute API key (x402c_...)
)

msg = client.messages.create(
    model="qwen-2.5-32b",
    max_tokens=512,
    messages=[{"role": "user", "content": "Hello from the Grid"}],
)
print(msg.content[0].text)
print(msg.usage)  # input_tokens, output_tokens, plus cost_usd

// npm i @anthropic-ai/sdk
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://grid.x402compute.cc",
  apiKey: process.env.SGL_API_KEY, // a Singularity Layer compute API key (x402c_...)
});

const msg = await client.messages.create({
  model: "qwen-2.5-32b",
  max_tokens: 512,
  messages: [{ role: "user", content: "Hello from the Grid" }],
});
console.log(msg.content);

Streaming

with client.messages.stream(
    model="qwen-2.5-32b",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Write a haiku about GPUs."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Endpoints

Authentication and payment

• x-api-key: x402c_... or Authorization: Bearer x402c_...: billed to the key owner’s credits.
• Authorization: Bearer <session token>: billed to the signed-in account.
• X-Payment: <x402 payload>: pay per request. Without auth, the 402 body is an Anthropic error (billing_error) that also carries the x402 fields (x402Version, accepts, price_usd), so x402 clients work unchanged.

What works

• model, messages, max_tokens, system (string or text blocks), stop_sequences, stream, temperature, tools, tool_choice (auto, any, tool, none).
• Content blocks: text, image (base64, vision models only, one image per request, non-streaming), tool_use, tool_result (text, with is_error).
• Stop reasons: end_turn, max_tokens, stop_sequence, tool_use, refusal.
• usage.input_tokens and usage.output_tokens. Cache token counts are always 0. usage.cost_usd is an extra field with the charge.
• Errors use the Anthropic shape ({"type":"error","error":{...},"request_id":...}) with a request-id header.

What does not work

• Extended thinking with a budget, structured outputs (output_config.format), strict tools.
• Server tools and Anthropic-defined tools: web search, web fetch, code execution, bash, text editor, computer use, memory, tool search. Define your own tool with an input_schema instead.
• document blocks (PDF or text), search results and citations.
• Image URLs and Files API images (send base64), images inside tool_result.
• mcp_servers, container, top_p, top_k, assistant prefill, max_tokens: 0, disable_parallel_tool_use: true.
• Any other unknown top-level field. The Grid names it in the error instead of ignoring it.

Good to know

• Context windows are small. Each model serves the context its nodes run, often 4k to 16k tokens. Check max_input_tokens in /v1/models. A prompt that leaves no room for a reply gets a 400 with the limit. A max_tokens above the room left is lowered to fit (the x-sgl-max-tokens-* headers report it), and output is capped at 8192 tokens.
• Long replies: stream them. A non-streaming reply that runs past the sync window returns 504 timeout_error with the job_id and x-should-retry: false, because a retry would start a second, separately billed job.
• Stop sequences are applied by the Grid, not the node. The reply is cut at the first match, but the node keeps generating until it stops, and those tokens are billed.
• Tool call input arrives as one input_json_delta per call, once the call is complete and valid JSON.
• Retries. The SDKs retry 429 and 5xx by default, which is safe: nothing is charged on those. Payment errors are sent with x-should-retry: false.
• Privacy. /v1/messages uses the API path where the Grid seals your prompt to the node. The browser path where your own client seals the prompt has no Anthropic equivalent; see Confidential inference.

Related