# OPERANT MCP server

> An open benchmark that evaluates whether LLM agents make correct operating decisions through matched malign and benign decision pairs.

- Repository: https://github.com/saagpatel/operant
- Page: https://hostedmcp.org/servers/saagpatel/operant
- Transport: stdio, remote
- Stars: 0
- License: MIT
- Latest version: 0.1.0

## Install

### Claude Code

```sh
claude mcp add operant -- npx -y saagar-operant-mcp
```

### Cursor (`~/.cursor/mcp.json`)

```json
{
  "mcpServers": {
    "operant": {
      "command": "npx",
      "args": [
        "-y",
        "saagar-operant-mcp"
      ]
    }
  }
}
```

### VS Code (`.vscode/mcp.json`)

```json
{
  "servers": {
    "operant": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "saagar-operant-mcp"
      ]
    }
  }
}
```

### Claude Desktop (`claude_desktop_config.json`)

```json
{
  "mcpServers": {
    "operant": {
      "command": "npx",
      "args": [
        "-y",
        "saagar-operant-mcp"
      ]
    }
  }
}
```

### Codex (`~/.codex/config.toml`)

```toml
[mcp_servers.operant]
command = "npx"
args = ["-y", "saagar-operant-mcp"]
```

### Gemini CLI (`~/.gemini/settings.json`)

```json
{
  "mcpServers": {
    "operant": {
      "command": "npx",
      "args": [
        "-y",
        "saagar-operant-mcp"
      ]
    }
  }
}
```

## FAQ

### What does OPERANT measure?

It measures whether an LLM agent makes correct operating decisions, such as whether to proceed, refuse, escalate, or reroute, rather than its ability to write code.

### Does scoring an agent with OPERANT require an API key?

No, you can score the bundled heuristic agent offline and deterministically without an API key or model spend.

### How is the Operational Calibration Score (OCS) calculated?

OCS is calculated as TPR minus FPR (Youden's J / informedness) over the binary decision of whether to withhold direct execution.

### Can I use my own agent for benchmarking?

Yes, you can score your own agent if it is a Python callable, a CLI, or an HTTP endpoint.