Developers · sparehash

Quickstart

Your First Request

Base URL: https://your-coordinator.fly.dev/v1

curl

curl https://your-coordinator.fly.dev/v1/chat/completions \
  -H "Authorization: Bearer $SPAREHASH_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-8b", "messages": [{"role": "user", "content": "Hello!"}]}'

Node.js · streaming

import OpenAI from 'openai';

const client = new OpenAI({ apiKey: process.env.SPAREHASH_KEY, baseURL: 'https://your-coordinator.fly.dev/v1' });
const stream = await client.chat.completions.create({
  model: 'llama-3.1-8b-instruct',
  stream: true,
  messages: [{ role: 'user', content: 'Write a haiku about idle GPUs.' }],
});
for await (const chunk of stream) process.stdout.write(chunk.choices[0]?.delta?.content ?? '');

Python · embeddings

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["SPAREHASH_KEY"], base_url="https://your-coordinator.fly.dev/v1")
vectors = client.embeddings.create(model="nomic-embed-text", input=["hello", "world"])

Privacy

Private By Default

Before a request leaves sparehash, the privacy filter swaps credentials and personal data for placeholders. The machine that runs the model only ever sees [EMAIL_1] or [SECRET_1]; the real values are put back into the answer before it reaches you.

What's masked

API keys and tokens: OpenAI, Anthropic, Google, AWS, GitHub, Stripe, Slack, Hugging Face, JWTs, private keys
Passwords and secrets written as key=value, and passwords inside connection strings
Email addresses, phone numbers, card numbers, US SSNs and IBANs

And around it

The sparehash app runs models in a sandbox and never writes prompts to disk or logs
Hidden test requests check every node around the clock; nodes that fail are quarantined, then banned
Add X-Trusted-Only: 1 to use only verified, long-standing nodes

Know the limits

Names, street addresses and secrets written in plain words ("my PIN is the year I was born") can't be recognised, so they aren't masked
The model has to read your text to answer it. Someone who takes their own computer apart could still read the parts that aren't masked
For health, legal or financial records, call the provider directly from your own account rather than through a shared network

Choose per request: standard (default), secrets, off

curl https://your-coordinator.fly.dev/v1/chat/completions \
  -H "Authorization: Bearer $SPAREHASH_KEY" \
  -H "X-Privacy-Filter: secrets" \
  -H "X-Trusted-Only: 1" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-32b", "messages": [{"role": "user", "content": "Draft a reply to [email protected]"}]}'
# x-sparehash-masked: number of values that were swapped out

Provider models

GPT, Claude and Gemini, Discounted

Sellers with spare capacity on official provider accounts list it on sparehash at a discount off the provider's list price. Call those models through the same key: put the provider in front of the model name. Requests go from sparehash straight to the provider's official API.

Official providers only

Provider	Model name
OpenAI	`openai/gpt-6.1-sol`
Anthropic	`anthropic/claude-sonnet-5.5`
Google Gemini	`google/gemini-3.8-flash`
Mistral AI	`mistral/mistral-small-2603`
xAI	`xai/grok-4.7`
DeepSeek	`deepseek/deepseek-v4.1-flash`
Alibaba Cloud (Qwen)	`alibaba/qwen3.8-flash`

How it works

Your request goes to the deepest discount that's live and healthy, and to the next seller if one fails before answering
You pay the provider's list price minus the seller's discount, from your USDC credits
Sellers never see requests, and the privacy filter applies here too
Tools, images and the provider's other features pass through untouched
What's on sale right now: GET /v1/models or the markets page

Node.js · same client as above

const reply = await client.chat.completions.create({
  model: 'anthropic/claude-sonnet-5.5',
  messages: [{ role: 'user', content: 'Summarise this contract in three bullet points.' }],
});
// x-sparehash-route: market · x-sparehash-discount-bps: 7500

Compatibility

What Works Today

Supported

POST /v1/chat/completions with stream and stream_options.include_usage
POST /v1/embeddings, float or base64
GET /v1/models, with pricing and license per model
max_tokens, temperature, top_p, stop, seed, response_format: json_object
On provider models (anthropic/…, openai/…): everything that provider supports

Not yet on sparehash nodes

Tool / function calling (returns a clear 400)
Image inputs, and n greater than 1
Fine-tuning, files, assistants and batch APIs

Errors

Status Codes

Status	Code	Meaning
400	`invalid_privacy_filter`	`X-Privacy-Filter` isn't `standard`, `secrets` or `off`
401	`invalid_api_key`	Missing, wrong or revoked key
402	`insufficient_quota`	Not enough credits for this request's worst case
404	`model_not_found`	Unknown model id, or no seller lists that provider model
429	`rate_limit_exceeded`	Requests or tokens per minute; honour `retry-after`
502	`upstream_failed`	Every node failed; you were not charged
503	`no_capacity`	No node or seller is serving that model right now

Request errors from a provider (e.g. an unsupported parameter) are passed through as the provider sent them.

Get Your Key

Sign in with your wallet, add USDC credits and create an API key.

Open the dashboard

OpenAI-Compatible, Open Models