# Model Runner

Stop copy-pasting boilerplate every time you need to call a different AI model.

![MCPize](https://img.shields.io/badge/MCPize-live-brightgreen) ![License: MIT](https://img.shields.io/badge/License-MIT-blue)

**Model Runner** is an MCP server that gives any AI assistant a unified interface to run completions, embeddings, image generation, and classification across all major providers. One tool call instead of per-provider API clients.

## Quick Start

Add to your `mcpServers` config:

```json
{
  "mcpServers": {
    "model-runner": {
      "url": "https://your-cloud-run-url/mcp"
    }
  }
}
```

Or run locally:

```bash
npm install
npm start
```

## Before / After

**Before:** Your assistant wants to classify customer feedback into sentiment buckets. It cannot call the OpenAI API natively, does not know the exact endpoint shape, and cannot handle auth headers.

```
// 30 lines of fetch boilerplate. Per provider. Per project.
// Auth headers, message array format, error shapes, all different.
```

**After:** One tool call:

```json
{
  "tool": "run_classification",
  "arguments": {
    "text": "Waited 40 minutes for support and got no answer",
    "labels": ["positive", "negative", "neutral"],
    "api_key": "sk-..."
  }
}
```

Output:

```json
{
  "label": "negative",
  "confidence": 0.97,
  "reasoning": "The customer experienced a long wait with no resolution, indicating a clearly negative experience."
}
```

## Tools

| Tool | What it does |
|------|--------------|
| `list_supported_models` | Browse the full catalog of models by provider and capability |
| `run_completion` | Text generation via OpenAI, Anthropic, Groq, or Mistral |
| `run_embedding` | Vector embeddings via OpenAI or Cohere |
| `run_image_generation` | Image generation via DALL-E 2 or DALL-E 3 |
| `estimate_tokens` | Estimate token count before making expensive API calls |
| `run_classification` | Zero-shot text classification with confidence scores |

## Who is this for?

- **AI assistant builders** who want their agent to invoke ML models without hardcoding provider-specific API logic into every project
- **Developers prototyping** who need a quick way to compare outputs across OpenAI, Anthropic, Groq, and Mistral without writing multiple API clients
- **Data teams running pipelines** who want a single MCP endpoint to classify, embed, or summarize records at scale without managing provider SDKs

## Health Check

Both endpoints return the same response and require no authentication:

```
GET /
GET /health
```

Response:

```json
{
  "status": "ok",
  "server": "model-runner",
  "version": "1.0.0",
  "tools": 6
}
```

## Built by

[Mastermind HQ](https://mastermindshq.business) - AI tools built for builders.

## License

MIT