ModelFit-MCP

Hardware-Aware Hugging Face Discovery, Memory Fitting, Empirical Benchmarking & Swappable Local Model Gateway for AI Agents.

ModelFit-MCP connects AI agents (Claude Desktop, Cursor, custom autonomous agents) to Hugging Face with built-in hardware awareness. It profiles host specs (CPU, RAM, GPU, VRAM), calculates exact model memory footprints, filters out models that would trigger CUDA Out of Memory (OOM) errors, and spins up a swappable local gateway so models can be changed dynamically without modifying application code.

Key Features

• Zero OOM Crashes: Automatically evaluates parameter count, precision (fp32, fp16, int8, int4), and 25% activation headroom before suggesting or loading models.
• Empirical Accuracy & Latency Benchmarking: Automatically runs test evaluations across candidate models that satisfy hardware specs to rank them by real-world inference speed and confidence.
• High-Speed Ensembles: Discovers ultra-lightweight models that consume $\le 35%$ of hardware headroom and aggregates them via weighted_average, majority_vote, or top_confidence strategies.
• Swappable Architecture: Your client code interacts with an abstract gateway (gateway.predict() or POST /predict). Models can be upgraded or hot-swapped without touching your application code.
• Memory Purging: Unloads old weights and triggers torch.cuda.empty_cache() on every swap to prevent VRAM memory leaks.
• Multi-Modal: Normalizes outputs across image-classification, object-detection, text-generation, and zero-shot-image-classification.
• Cross-Language Ready: First-class support for Python, Flutter/Dart, Node.js, and cURL.

System Architecture

flowchart TD
    User["Your App / AI Agent\n(Plant App, Flutter, Web, CLI)"] -->|predict / ensemble| Gateway["Local Model Gateway\n(In-Process or http://127.0.0.1:7860)"]
    
    subgraph Engine["ModelFit-MCP Engine"]
        Gateway --> ActiveModel["Active Model(s)\n(MobileNet / ViT / ResNet)"]
        MCP["MCP Server / CLI\n(modelfit)"] -.->|Hot Swap + Memory Purge| ActiveModel
        Profiler["hardware.py\n(VRAM & RAM Profiler)"] --> HFFilter["hf_client.py\n(Sizing & Ranking)"]
        HFFilter --> Evaluator["evaluator.py\n(Benchmarking & Latency)"]
        HFFilter --> Ensemble["ensemble.py\n(Multi-Model Aggregator)"]
        Evaluator --> MCP
        Ensemble --> MCP
    end

Quickstart

1. Installation

git clone https://github.com/DumboDhruvi/ModelFit-MCP.git
cd ModelFit-MCP
pip install -e .

2. CLI Usage

Inspect your hardware headroom:

modelfit specs

Search Hugging Face models guaranteed to fit your machine:

modelfit search "plant disease" --task image-classification

Benchmark multiple candidate models on your hardware:

modelfit benchmark "nateraw/food,google/vit-base-patch16-224" --samples "sample1.jpg,sample2.jpg"

Run ensemble inference across lightweight models:

modelfit ensemble "model-a,model-b" --input "sample.jpg" --strategy weighted_average

Start the local micro-API daemon:

modelfit serve --port 7860

Hot-swap models on the fly:

modelfit swap "google/vit-base-patch16-224" --task image-classification

MCP Server Tools (Claude Desktop & Cursor)

Add ModelFit to your claude_desktop_config.json:

{
  "mcpServers": {
    "modelfit": {
      "command": "modelfit-server"
    }
  }
}

Available MCP Tools:

Integration Modes

Mode A: In-Process Python Adapter (Zero Latency)

from modelfit.adapter import gateway

# 1. Load initial model
gateway.load_model("linkanjarad/mobilenet_v2_1.0_224-plant-disease-identification")

# 2. Abstract prediction
results = gateway.predict("leaf.jpg")
print(results[0].label, results[0].score)

# 3. Hot-swap later with ZERO code changes below
gateway.load_model("google/vit-base-patch16-224")
results = gateway.predict("leaf.jpg")

Mode B: High-Speed Ensemble Inference

from modelfit.ensemble import EnsembleGateway

ensemble = EnsembleGateway()
candidates = ensemble.find_ensemble_candidates("plant disease", max_models=3)
model_ids = [m["model_id"] for m in candidates]

predictions = ensemble.predict_ensemble(
    input_data="leaf.jpg",
    model_ids=model_ids,
    strategy="weighted_average"
)

for p in predictions:
    print(f"{p.label}: {p.score:.2f} (votes: {p.votes})")

Mode C: Local Micro-API (Flutter, Node.js, Web, cURL)

Start the background daemon:

modelfit serve --port 7860

Query or swap over HTTP:

# Predict
curl -X POST http://127.0.0.1:7860/predict \
  -H "Content-Type: application/json" \
  -d '{"input": "leaf_sample.jpg"}'

# Hot-Swap
curl -X POST http://127.0.0.1:7860/swap \
  -H "Content-Type: application/json" \
  -d '{"model_id": "google/vit-base-patch16-224", "task": "image-classification"}'

See examples/flutter_integration_example.dart for a complete Flutter service.

Testing

Run the full test suite:

python3 -m unittest discover tests -v

All tests complete in under 2 seconds.

License

MIT License. See LICENSE for details.