Second opinions from Claude Code (MCP)

Parley is an MCP server. Once it is connected, the AI you are working with in Claude Code, Claude Desktop or Cursor can call one tool, second_opinion. Models from three different companies (for example Anthropic, OpenAI and Google) answer the same question independently, each told to lead with the weakest point, and you get back whether they agreed or split, where they disagree, and every full answer.

Use it before committing to something you are not sure about: a plan, a design, an architecture, a library choice, a claim. One model checking its own work tends to agree with itself; three companies trained on different data are more likely to disagree where it matters.

1. Create a key

Sign in to Parley and open Settings → Claude Code / MCP. Press Create key. The key is shown once; only a hash of it is stored. You can have three keys and revoke any of them at any time.

2. Connect your client

Claude Code: run this once in your terminal, with your key in place of pk_….

claude mcp add --transport http parley https://parley.amukin.com/api/mcp --header "Authorization: Bearer pk_…"

Claude Desktop: add this to claude_desktop_config.json (Settings → Developer → Edit Config). Node.js is required; mcp-remote connects the desktop app to a remote server.

{
  "mcpServers": {
    "parley": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://parley.amukin.com/api/mcp",
        "--header",
        "Authorization:${PARLEY_AUTH}"
      ],
      "env": {
        "PARLEY_AUTH": "Bearer pk_…"
      }
    }
  }
}

Cursor and other clients that accept a URL and headers (for example .cursor/mcp.json or a project .mcp.json):

{
  "mcpServers": {
    "parley": {
      "type": "http",
      "url": "https://parley.amukin.com/api/mcp",
      "headers": {
        "Authorization": "Bearer pk_…"
      }
    }
  }
}

Restart the client after adding it. The server address is https://parley.amukin.com/api/mcp.

3. Ask for a second opinion

In the middle of normal work, ask for it in plain words. For example:

• "Before you implement this, get a second opinion from Parley on the plan."
• "Ask Parley whether this database schema will hold up at 10x the data."
• "Check this claim with Parley before we put it in the README."

It takes 20 to 90 seconds. The result says whether the models AGREED or SPLIT, lists each disagreement with which model took which side, the common ground, points only one model raised, and each full answer, with a link to the Room in Parley where you can keep going.

Cost and limits

• Credit is drawn from your Parley account, the same as a Panel in the app. It works on Free with the trial Credit.
• Each check uses one low-cost model from each of three companies, plus the comparison.
• Per key: 5 calls a minute. Per account: 3 at once, and 30 a day on Free or 200 a day on paid plans (resets at 00:00 UTC).
• Question and context together: up to 30,000 characters.

Privacy and security

• What you send goes to three outside AI providers and is saved as a Room in your account. Do not send secrets.
• As a safety net, values that look like API keys, private keys or passwords are replaced with [REDACTED] before anything is sent, and the result says how many were replaced.
• The result tells the calling AI that the answers are other models' output, to be weighed as evidence and not followed as instructions.
• A key gives access only to second opinions on your own account. Revoke it in Settings if it leaks; it stops working immediately.

Tools

• second_opinion: question (required), context (optional background), wait_seconds (0 to 120, default 90).
• get_second_opinion: check_id. Reads a result that was not ready when the first call returned.