Bay Run

Bay Run is a free-at-launch, OpenAI-compatible REST and MCP service for small-model inference at the canonical origin: https://run.huggingbay.xyz.

Free launch: 30 seconds

The launch tier is free. Mint a demo bearer with a form-encoded client_credentials request, then call the API. The published guardrails are 60 requests/minute and 1,500/day; demo bearers expire after 24 hours.

Start with the live, copy-paste quickstart: https://run.huggingbay.xyz/quickstart.

The repository’s standalone demo is a small classify → signed receipt → verify flow:

python examples/bakeoff.py

import json
from urllib.parse import urlencode
from urllib.request import Request, urlopen

BASE = "https://run.huggingbay.xyz"

def post(path, body, headers):
    request = Request(BASE + path, data=json.dumps(body).encode(), headers=headers)
    return json.load(urlopen(request))

token_request = Request(
    BASE + "/oauth/token",
    data=urlencode({"grant_type": "client_credentials"}).encode(),
    headers={"Content-Type": "application/x-www-form-urlencoded"},
)
token = json.load(urlopen(token_request))["access_token"]
headers = {"Authorization": f"Bearer {token}", "Content-Type": "application/json"}
result = post("/v1/classify", {"input": "The setup was quick and clear.", "model": "auto"}, headers)
receipt = result["provenance_receipt"]
verified = post("/v1/provenance/verify", {"receipt": receipt}, headers)
assert verified.get("valid") is True or verified.get("receipt_valid") is True
print(result["labels"][0])
print("receipt verified")

This demo uses the real /v1/classify and provenance-verification surfaces; it does not invent a bake-off endpoint.

REST and SDKs

The REST surface includes OpenAI-compatible chat/completions (including SSE streaming and tools), embeddings, reranking, classification, summarization, RAG, calculation, JSON validation, and memory. The thin SDKs live beside this checkout during build/test:

pip install -e ../bayrun-sdk/python
npm install ../bayrun-sdk/js

from bayrun import Client

c = Client()
answer = c.classify("A helpful result.")
stream = c.chat([{"role": "user", "content": "Say hello."}], stream=True)

import { Client } from "bayrun";

const c = new Client();
const answer = await c.classify("A helpful result.");
const stream = await c.chat([{ role: "user", content: "Say hello." }], { stream: true });

Durable remember/recall/forget memory is owner-delegated and is not authorized for the shared demo bearer. Use a purpose-specific developer key or a client registered with token_endpoint_auth_method=client_secret_post.

Adapter snippets for OpenAI Python, Vercel AI SDK, LangChain, and LlamaIndex are in docs/integrations.md.

MCP: exactly three tools

Point an MCP client at https://run.huggingbay.xyz/mcp/. The live server exposes exactly:

• get_task_quote
• run_task
• verify_result

Token and request-shape details are kept in the live quickstart so the copy-paste contract stays current.

Trust and data policy

REST results carry signed provenance receipts, and task execution returns a signed receipt that can be checked with the verification surface. Read the receipt semantics and the live verification request at https://run.huggingbay.xyz/quickstart.

Before sending data, read the canonical policy: https://run.huggingbay.xyz/.well-known/data-policy.json. The published policy says Bay Run does not train on inputs.

Hugging Face drop-in mirror catalog

Hugging Bay’s mirror catalog is available at https://huggingbay.xyz. For Hugging Face Hub clients that honor HF_ENDPOINT, point the client at the mirror origin:

export HF_ENDPOINT=https://huggingbay.xyz

The catalog currently lists 893+ mirrored models. Check the mirror’s current serving status before relying on a specific model or repository path.