Weighthouse
Use cases / MCP and the API
Give your agents access to your model library.
Finding a model is only part of the job. You still need to check what you already have, download the files you want, and keep track of what's on disk once you’ve downloaded them.
Weighthouse gives your agents access to your managed library. Through its built-in MCP server and API, your agents can search Hugging Face, inspect local copies of your models, queue models up for download, check your models for updates, and verify the files you already have on disk.
Give each of your agents or automation tools a scoped key with only the permissions that it needs. Use Weighthouse interactively with an assistant, or integrate your model library into your own scripts and automated workflows. You decide what operations each key allows.
Start with the models you already have
Before downloading more models, your agent can check your existing library. It can see which models are available locally, where they’re stored, which revisions you have, and what files belong to each managed copy.
That gives you a useful starting point for questions like “Do I already have this model?” or “Which quants did I download?” Your agent can just inspect your library and tell you, instead of asking you to find a directory and paste its contents into a conversation.
When you do need something new, your agent can search Hugging Face, plan a download, and add the files you want to the same managed download queue that you use in the web interface.
Explore your options before downloading
Your agent can also help you research models on Hugging Face. It can search by keyword or task, sort results by recent updates, downloads, likes, or trending activity, and inspect the files available in a repository.
That lets you ask questions like “Which quantizations are available?” or “How large are the files I’d need to download?” Your agent can check the repository’s file list and sizes without downloading the model weights first.
Once you’ve picked a model and the files you want, your agent can plan the download and add it to your queue, provided its key has permission.
Give each agent and tool the access it needs
Administrators can create separate keys for your agents and automation tools. Each key can have up to four scopes: searching Hugging Face, reading your library, queueing downloads and updates, or changing stored content.
You can give an assistant permission to search and inspect your models, then add download permissions when you want it to fetch files for you. Operations like moving and deleting stored models require the appropriate scope.
Gated repositories use the Hugging Face token configured on your Weighthouse server. You need to get access to the gated repository through Hugging Face itself. Once your account has access, your agents can request downloads through Weighthouse using their own scoped keys. Your Hugging Face token stays on the server.
Requests made through each key are recorded in the key’s request history. You can review that history and revoke a key whenever you want to stop its access.
Decide how much you want to automate
Sometimes, you just want to ask your assistant, “What’s in my library?”, or have it download a model you’ve already picked. You can work interactively, reviewing what it finds and telling it what to do next.
You can also use MCP and the API in your own automated workflows. A script could check for upstream changes on a schedule, or an agent could queue downloads as part of preparing a workload. If you want updates to happen without reviewing each one yourself, your automation can request them using a key with the necessary permissions.