Hugging Bay for VS Code and Cursor

Browse models, estimate RAM fit, download with a hash check and use local tools.

Version 0.2.0 preview. Off by default. Requires VS Code 1.105 or compatible Cursor, with Node 22.12 or later in the extension host. Runs on the editor's local UI host, including when a workspace is remote.

Install and enable

1. Install the VSIX with Extensions: Install from VSIX.
2. Set these in user settings:

{
  "huggingBay.enabled": true,
  "huggingBay.v2Enabled": true,
  "huggingBay.localOperations": false,
  "huggingBay.mcpCommand": "hugging-bay-mcp",
  "huggingBay.mcpArgs": ["--profile", "full"]
}

All three flags default to false. The v2 commands stay hidden until both enabled and v2Enabled are true. Downloads, app handoff, Ollama and settings edits also require localOperations. These are application settings. Workspace settings cannot enable the integration or choose executable paths. Untrusted and virtual workspaces are unsupported.

Browse and download

Choose Hugging Bay: Browse Models from the command palette. Search, select a model and review its fit result in a JSON editor. Then choose an action.

Search and fit use the public REST API. They do not need a local MCP server. The extension sends the local host's total RAM as an estimate, with its platform and architecture shown in the result. Other apps, GPU memory and context length can limit a run. RAM fit does not prove that a model will run.

Download with hash check lets you choose an exact file and destination. It rechecks the download plan, streams the admitted file, checks SHA-256 and size, and records completion before saving. Existing files and symlinks are preserved. Partial files are removed on failure. No model weights are buffered in memory.

Hugging Bay receives queries and RAM budgets; model execution stays local.

Open in the app or run with Ollama

Hugging Bay: Open in Hugging Bay selects a downloadable GGUF file and opens hbay://chat/<model-id>?repo=<repo>&file=<file>. Install an app version that handles this link first. The app checks its own download and run requirements.

Hugging Bay: Run with Ollama selects a file and a supported registry tag, checks the manifest hash and exact model layer against the download plan, then pulls with the local Ollama API. Ollama checks the downloaded layers. The extension checks the resulting local manifest digest before opening an Ollama chat in a terminal. Model names are passed as executable arguments.

Install Ollama separately. Its default API is http://127.0.0.1:11434. Override huggingBay.ollamaUrl with another local HTTP origin and huggingBay.ollamaCommand with the installed executable path if needed. The API host and terminal use the same Ollama server.

The Hugging Bay registry is a separate server preview controlled by HUGGING_BAY_OLLAMA_REGISTRY, off by default. A 404 can mean that it is off or the tag is unsupported. The current template adapter supports Qwen2.5 text chat. There is no automatic fallback to another registry.

Add MCP to agent settings

Install the Hugging Bay MCP server separately. Choose Hugging Bay: Add MCP to Agent Settings, select the settings format, then choose your existing user settings file:

• VS Code mcp.json uses a servers object with a stdio entry. Choose the user mcp.json, not VS Code settings.json.
• Cursor or Claude Code MCP JSON uses a mcpServers object. Choose the user's Cursor mcp.json or Claude Code .claude.json.

The command inserts a hugging-bay entry using mcpCommand and mcpArgs. It preserves comments, unrelated settings and other servers, and refuses to replace an existing hugging-bay entry. Review and save the open document. The extension does not start an agent or install a server. Your agent controls which MCP tools it permits.

For a source checkout, set mcpCommand to node and mcpArgs to ["/absolute/repo/packages/mcp/index.mjs", "--profile", "full"].

Existing MCP commands

The original commands remain under huggingBay.enabled:

• Find a Model for This Machine asks for a task and uses MCP find_runnable.
• Download Verified asks for an artifact id and uses MCP get_download_plan plus the same verified REST download client. Requires localOperations.
• Run Locally uses MCP run_local. Requires localOperations, the separate hbay runtime, and HBAY_ENABLE_RUN=1 in the editor launch environment. An MCP release without this tool reports an error. The extension owns the MCP session; reloading or changing integration settings closes it.

Build and check

Follow the source setup in the parent guide. From this folder:

npm test
npm run test:bundle
npx --no-install vsce package --no-dependencies --out /tmp/hugging-bay-0.2.0.vsix

The build bundles runtime dependencies and their license notices. VSIX packaging excludes source, tests and local dependencies. Tests use a synthetic VS Code API, local MCP protocol fixture and HTTP byte fixtures. They cover flags, browsing, fit, downloads, hash failures, app links, Ollama and settings preservation. They do not download weights or run a model. No publishing is part of these steps.

Build from this repository with npm ci and npm run build in this directory. The lockfile includes the sibling integrations package, so a registry publication is not required.