# chatgpt-interlocutor

An MCP server that gives Claude Code a ChatGPT second opinion. Cross-model adversarial review without leaving your terminal.

## Why

When you're working in Claude Code, sometimes you want a second perspective — a different model's take on the same problem. Not because one is better, but because disagreement surfaces blind spots.

This server adds two tools to every Claude Code session:

- **`ask_chatgpt`** — Send a question to ChatGPT and get its response inline. Second opinions, alternative approaches, sanity checks.
- **`compare_approaches`** — Send the same prompt to both models and get a structured comparison. Specificity, depth, accuracy, voice.

## Install

### Prerequisites

You need an [OpenAI API key](https://platform.openai.com/api-keys).

### Add to Claude Code

Add this to your `~/.claude.json` under `mcpServers`:

```json
{
  "mcpServers": {
    "chatgpt-interlocutor": {
      "command": "npx",
      "args": ["-y", "chatgpt-interlocutor"],
      "env": {
        "OPENAI_API_KEY": "sk-your-key-here"
      }
    }
  }
}
```

Restart Claude Code. You should see `ask_chatgpt` and `compare_approaches` in your available tools.

### Or install globally

```bash
npm install -g chatgpt-interlocutor
```

Then configure with the direct path:

```json
{
  "mcpServers": {
    "chatgpt-interlocutor": {
      "command": "chatgpt-interlocutor",
      "env": {
        "OPENAI_API_KEY": "sk-your-key-here"
      }
    }
  }
}
```

## Tools

### `ask_chatgpt`

Get a second opinion from ChatGPT.

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `prompt` | string | *required* | The question or task to send |
| `model` | string | `gpt-4o-mini` | `gpt-4o-mini`, `gpt-4o`, or `o3-mini` |
| `system` | string | — | Optional system prompt for context |
| `temperature` | number | `0.7` | Creativity 0–2. Lower = more focused |

### `compare_approaches`

Benchmark Claude's output against ChatGPT's on the same task.

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `task_description` | string | *required* | What both models are being asked to do |
| `prompt` | string | *required* | The prompt to send to ChatGPT |
| `your_output` | string | — | Claude's output for side-by-side comparison |
| `model` | string | `gpt-4o-mini` | Which ChatGPT model to use |

When `your_output` is provided, the response includes a comparison framework:
- **Specificity** — Which is more concrete and actionable?
- **Depth** — Which goes deeper into the problem?
- **Voice** — Which sounds more human / less generic AI?
- **Structure** — Which is better organized?
- **Accuracy** — Which is more correct?
- **Differentiation** — Where do they diverge most?

## Models

| Model | Cost | Best for |
|-------|------|----------|
| `gpt-4o-mini` | Cheapest | Quick second opinions, brainstorming |
| `gpt-4o` | Mid | Deeper analysis, complex reasoning |
| `o3-mini` | Mid | Math, logic, structured reasoning |

## How it works

This is a [Model Context Protocol](https://modelcontextprotocol.io/) server that runs over stdio. Claude Code launches it as a subprocess, sends tool calls, and gets responses back. Your prompts go to the OpenAI API using your own key — nothing is stored or proxied.

## License

MIT