gguf-mcp

An MCP server that inspects local model files — GGUF and safetensors — so Claude and other LLMs can answer questions about the models on your disk:

• "What is this .gguf? Architecture, quantization, parameter count?"
• "Will this model fit in my 12 GB GPU at 8k context?"
• "What tensors are inside, with what shapes?"
• "Show me its chat template / RoPE settings / tokenizer config."

Headers only. The parser never touches tensor data, so inspecting a 70 GB model takes milliseconds and a few MiB of I/O. No network, no API keys, no telemetry — your files never leave your machine.

Quick start

Claude Code

claude mcp add gguf -- npx -y gguf-mcp

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "gguf": {
      "command": "npx",
      "args": ["-y", "gguf-mcp"]
    }
  }
}

The same npx invocation works in Cursor, Windsurf, and any other MCP client.

Tools

| Tool | What it does | |------|--------------| | inspect_model | One-call summary: format, architecture, parameters, quantization, context length, file size, tensor count | | list_tensors | Tensor names, shapes, and storage types — filterable (attn, blk.0, ...) | | estimate_vram | Fit check: exact weights size + modeled fp16 KV cache for your chosen context length | | get_metadata | The GGUF key-value store (or safetensors __metadata__), filterable by key |

Paths can be a .gguf file, a .safetensors file, a *.safetensors.index.json, or a model directory (sharded HuggingFace layouts are aggregated across shards). Extension-less GGUF blobs — like the ones in Ollama's ~/.ollama/models/blobs — are detected by magic bytes.

Design notes

• Context-friendly by construction. A tokenizer vocabulary is 100k+ strings; metadata arrays are returned as {count, sample} summaries and long strings (chat templates) are truncated with a marker. The full data stays on disk where it belongs.
• Honest estimates. estimate_vram reports exact on-disk weight bytes plus the standard KV-cache formula (2 × layers × context × KV heads × head dim × 2 bytes), and says what it excludes rather than faking precision.
• Defensive parsing. Magic checks, version checks (incl. big-endian detection), truncation detection, and sanity caps on header sizes — malformed files produce specific, actionable errors.
• Zero runtime dependencies beyond the MCP SDK and zod. The GGUF binary reader and safetensors parser are hand-rolled and unit-tested against synthetic files built in the test suite — no fixtures, no downloads.

Development

npm install
npm test                 # offline unit tests (vitest) — synthetic model files
npm run build            # tsc → dist/
node scripts/smoke.mjs   # end-to-end: generates models, drives the server over stdio

Architecture: src/gguf.ts (binary header parser + VRAM math) and src/safetensors.ts (JSON header + shard index) are pure logic with no MCP imports; src/index.ts is the MCP wiring and path/format detection.

Out of scope

Tensor statistics (would require reading data), PyTorch .bin (pickle — unsafe by design), ONNX, and remote HuggingFace queries (HuggingFace has an official MCP server for that).

License

MIT