ask-james 0.0.4

Model Context Protocol server that offers a 'James' second-opinion review tool.

ask-james

Tired of copy/pasting between assistants just to get a second opinion?

ask-james ends that dance. It is an open-source Model Context Protocol (MCP) server that wraps your “ask another assistant what they think” habit into a single tool – james.review. James behaves like a trusted colleague who reviews designs or implementation plans, and because the server runs over stdio any MCP-compatible host (Claude desktop, Kiro, Cursor/Windsurf, custom orchestrators) can invoke him on demand.

Instead of bouncing between tabs, James gives you:

• one command (ask-james) to plug into any assistant
• a consistent persona that critiques, not rewrites
• flexible model backing via LiteLLM, so the review voice can be Claude today and GPT tomorrow

Highlights

• Single-focus tool – james.review ingests a proposal plus contextual metadata and returns a JSON critique (summary, risks, questions to ask, positives, and next steps).
• LLM-agnostic – Uses LiteLLM so you can point James at any supported provider/model just by changing environment variables.
• Named persona – James never overwrites the primary assistant. He only critiques, surfaces assumptions, and suggests safer next steps.
• Stdio transport – Works out of the box with MCP clients such as the Claude desktop app or custom hosts.
• Human-in-the-loop – James never takes action; he just challenges the suggestion so you or the host assistant can choose the best final answer.

Why this exists

Do you ever:

• ask your main assistant for a proposal,
• copy the whole thing into another assistant to get a “different take,” and
• paste the critique back into the first chat to reconcile it all manually?

That workflow is powerful but clunky. ask-james automates it without changing the way you think: you still own the conversation, you just say “ask James what he thinks,” and the second-opinion pass happens behind the scenes. James shows up with structured critique, assumptions to verify, and questions to ask—no tab juggling required.

Best part: it works in both directions. If you bounce between, say, Claude and Kiro, install ask-james in both. When you’re in Claude you can ask Kiro-for-James’s-take, and when you’re in Kiro you can pull Claude’s tone through the same tool. The review pattern stays consistent no matter which assistant you start with.

When to use James

ask-james shines whenever you need to challenge an assistant’s suggestion before acting:

• Implementation plans – Your primary assistant proposes an architecture or refactor; James critiques it so you don’t blindly follow the first answer.
• Product or policy decisions – Use James to stress-test assumptions, hidden risks, or compliance angles before presenting a recommendation.
• Support / operations workflows – When an assistant suggests a customer-facing response, have James flag risky language or missing steps before sending it.
• High-stakes automation – Anytime acting on an LLM’s plan has real impact (deploys, refunds, irreversible operations), call James so the decision-maker (you or the host model) hears dissenting views.

James never takes the decision away: he is explicitly a reviewer. You (or the primary assistant) remain responsible for the final call, with James providing the structured critique that normally requires a second human.

Quick start (step-by-step)

1. Reference James via uvx

   You don’t need to install anything manually. Just point your MCP host at:

   {
     "name": "Ask James",
     "command": "uvx",
    "args": ["ask-james"],
     "env": {
       "OPENAI_API_KEY": "sk-your-openai-key",
       "ASK_JAMES_MODEL": "gpt-5.2"
     }
   }
   

   (uvx downloads and caches the package automatically.)

2. Swap in your provider’s key/model

   LiteLLM reads provider-specific environment variables (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY). Embed the correct pair for the provider you’re using directly in the MCP config. Example for OpenAI:

   {
     "name": "Ask James",
     "command": "uvx",
    "args": ["ask-james"],
     "env": {
       "OPENAI_API_KEY": "sk-your-openai-key",
       "ASK_JAMES_MODEL": "gpt-5.2"
     }
   }
   

   Swap OPENAI_API_KEY for whichever variable LiteLLM expects (Anthropic, Google, Azure, Ollama, etc.).

3. Save the config in your MCP host

   Paste the JSON into the host’s MCP configuration file (claude_desktop_config.json, mcp-tools.json, etc.). Once saved, the host automatically launches James via uvx whenever you ask for a second opinion.

Configuration

James relies on LiteLLM’s environment variables. Set your provider’s API key using the variable name LiteLLM expects (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY) plus ASK_JAMES_MODEL. The tool schema (including optional fields like model or temperature) is defined in ask_james/server.py if you need to inspect it.

Using with LLM-based assistants

All hosts need the same three things:

1. An MCP config entry pointing to ask-james.
2. The LiteLLM credentials embedded in that config via an env block.

Use this template everywhere (swap the key name for your provider; command stays uvx so the host can fetch/run the package automatically):

{
  "name": "Ask James",
  "command": "uvx",
  "args": ["ask-james"],
  "env": {
    "OPENAI_API_KEY": "sk-your-openai-key",
    "ASK_JAMES_MODEL": "gpt-5.2"
  }
}

Claude Desktop (claude_desktop_config.json)

{
  "tools": [
    {
      "name": "Ask James",
      "command": "uvx",
      "args": ["ask-james"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key",
        "ASK_JAMES_MODEL": "gpt-5.2"
      }
    }
  ]
}

Kiro (Settings → MCP Tools)

{
  "id": "ask-james",
  "name": "Ask James",
  "command": "uvx",
  "args": ["ask-james"],
  "env": {
    "OPENAI_API_KEY": "sk-your-openai-key",
    "ASK_JAMES_MODEL": "gpt-5.2"
  }
}

Cursor / Windsurf (mcp-tools.json)

{
  "tools": [
    {
      "name": "Ask James",
      "command": "uvx",
      "args": ["ask-james"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key",
        "ASK_JAMES_MODEL": "gpt-5.2"
      }
    }
  ]
}

Tool hints (how to talk to James)

James listens for whatever the primary assistant passes as the proposal and optional context. Here are a few prompts you can give your assistant (Claude, Kiro, Cursor, etc.) so it knows when to call james.review:

• “Draft an MCP server plan, then ask James to critique it.”
• “Here’s an implementation plan—call James for a second opinion before finalizing.”
• “I’m about to send this customer response; ask James to flag any risks.”
• “Challenge this architecture with James and summarize what he worries about.”

Typical output James sends back (summarized in plain language):

• Summary – Does the plan look solid overall? Should you proceed, tweak, or rethink?
• Key concerns – Explicit issues with severity, impact, and suggestions.
• Questions – Follow-ups James would ask the author.
• Positive signals – Parts that look good so you know what not to over-optimize.
• Next steps & assumptions – Actionable follow-ups and things to verify.

If you need to integrate James programmatically, the james.review tool accepts JSON fields like proposal, context, review_focus, constraints, assumptions, risk_profile, max_questions, model, and temperature; and it returns the structured fields shown above plus model_used and raw_response. Most hosts serialize these automatically, but the schema is available in ask_james/server.py if you need to inspect it directly.

Recommendation / reasoning

• MCP stdio mode makes this trivial to embed into Claude, VS Code MCP hosts, or custom orchestrators without running a network service.
• LiteLLM keeps you model-neutral and instantly compatible with OpenAI, Anthropic, Google, Azure OpenAI, Ollama, etc. Changing James's voice is as simple as pointing at a different model alias.
• Persona-first design: James critiques, flags risks, and asks questions; the host model remains responsible for the final answer. This avoids “two agents arguing” and mirrors human peer review.

Development

• Code lives under ask_james/. server.py houses the tool definition, prompt builder, and LiteLLM call.
• Install dev dependencies with pip install -e .[dev] (add your preferred extras, e.g., ruff, pytest).
• Run python -m ask_james for manual stdio testing.
• Tests are not bundled yet; contributions welcome!

Contributing

Issues and pull requests are welcome. If you’re proposing substantial changes (new tools, new personas), please open an issue to discuss the use case so the “ask James” UX stays focused.

License

Released under the MIT License. See LICENSE for details.