Skip to main content
The Hugging BayFind a model

AI models that run on this device.

Two small chat models run in a computer’s browser here; every other model is a file you download.

Connect Hugging Bay

Add Hugging Bay's bounded open-model metadata tools to Claude, Cursor, Cline, OpenClaw, ChatGPT, or another MCP/OpenAPI-capable assistant.

Slash commands

Name the MCP server hugging-bay so commands read /mcp__hugging-bay__find-model-for-my-machine. List the seven commands with prompts/list; choose one and fill its arguments. Model names, memory and choices have completions. Commands guide tool use; request commands ask for your OK before submitting.

MCP (Claude Desktop / Claude Code / Cursor / Cline / OpenClaw)

{
  "mcpServers": {
    "hugging-bay": {
      "url": "https://huggingbay.xyz/api/mcp"
    },
    "bay-run": {
      "url": "https://run.huggingbay.xyz/mcp/"
    }
  }
}

hugging-bay supplies bounded public catalog metadata and evidence. bay-run supplies the separate specialist coprocessor; it does not proxy through Hugging Bay. Bay Run tool calls require its resource-bound OAuth bearer: compatible MCP clients follow the endpoint's protected-resource metadata automatically; otherwise mint the free MCP token through the Bay Run quickstart. Keep that token out of shared configuration files. The default Hugging Bay discovery profile exposes 10 bounded metadata tools (including the ready-to-run deployment_plan). The separate public verify profile exposes 17 read-only verification tools; publisher/full remain authenticated, and full-profile access is never inferred from the presence of an Authorization header. Test: curl -s https://huggingbay.xyz/api/mcp -H 'Content-Type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

OpenAPI (ChatGPT Actions / custom GPTs / any HTTP agent)

Import the REST contract from https://huggingbay.xyz/openapi.json. Start with /api/agents/find-model, /api/agents/recommend-prose, /api/v1/artifacts/{id}/report.

Hosted-file metadata

Inspect explicit hosted-file rows through the bounded file inventory API.

Hugging Face mirror

Only hosted files resolve; others return 404. Clients do not automatically retry Hugging Face. To use Hugging Face for an unhosted repo, unset HF_ENDPOINT and rerun the command in a new process. Transformers needs all required config, tokenizer and weights files. Use hf download on current clients; huggingface-cli is for older installs. File availability does not prove a model will load.

Public metadata can lag privacy changes by about 5 minutes; every download is checked again.

HF_ENDPOINT=https://hf.huggingbay.xyz hf download <owner>/<name> <file>
HF_ENDPOINT=https://hf.huggingbay.xyz python3 -c 'from huggingface_hub import hf_hub_download; hf_hub_download("<owner>/<name>", "<file>")'
HF_ENDPOINT=https://hf.huggingbay.xyz huggingface-cli download Qwen/Qwen2.5-0.5B-Instruct-GGUF
HF_ENDPOINT=https://hf.huggingbay.xyz python3 -c 'from transformers import AutoModel; AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2", trust_remote_code=False)'
HF_ENDPOINT=https://hf.huggingbay.xyz llama-cli -hf <owner>/<name>:<quant>

Hugging Face download guide

Machine-readable guides

We use an optional cookie to count visits.

Privacy