Getting Started

Getting Value Fast · Chapter 1 of 2 · Prerequisites: none · ~2 min read

In five minutes you will have a navigable semantic network of 19K+ ML models running locally, queryable through any MCP-compatible client.

---

Problem Space

HuggingFace Hub is a file host with search. You can filter by task, sort by downloads, and grep model cards. But you cannot ask "what's similar to this model but smaller and more code-focused?" — that requires structural understanding that HuggingFace doesn't encode.

ModelAtlas adds that structure. The pre-built network contains positions across 8 semantic dimensions, 170 anchor labels, and explicit model-to-model links — all queryable through a local MCP server.

---

Install

git clone https://github.com/rohanvinaik/ModelAtlas.git
cd ModelAtlas
uv sync

Download the Pre-Built Network

The semantic network is a ~~80MB SQLite file distributed via GitHub Releases:

mkdir -p ~/.cache/model-atlas
curl -L -o ~/.cache/model-atlas/network.db \
  https://github.com/rohanvinaik/ModelAtlas/releases/latest/download/network.db

Without this file, ModelAtlas starts with an empty database. You can build your own via hf_build_index, but the pre-built network includes multi-tier extraction that took days of distributed compute to produce.

Configure Your MCP Client

Add ModelAtlas to your client's MCP configuration:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "model-atlas": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/ModelAtlas", "model-atlas"]
    }
  }
}

Claude Code (.mcp.json in your project root):

{
  "mcpServers": {
    "model-atlas": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/ModelAtlas", "model-atlas"]
    }
  }
}

Replace /path/to/ModelAtlas with the actual path to your clone.

Your First Query

Once configured, ask your LLM client naturally:

"Find me a small code model with instruction-following that runs on consumer hardware"

The LLM will call navigate_models with decomposed parameters. ModelAtlas returns scored results with explanations of why each model matched — which anchors contributed, how each bank aligned.

See Query Examples for more of what you can ask.

---

What This Is Not

• Not a HuggingFace replacement. HF is a data source. ModelAtlas adds the structural layer HF doesn't expose.
• Not an API you call directly. The MCP server is the interface; the LLM is the user-facing layer.
• Not a cloud service. Everything runs locally. The database is a file on your disk.

---

Related Concepts

• Query Examples — cookbook of real queries
• The Gap — why this tool exists
• Data Distribution — versioning and update policy

---

← Home · Query Examples →