model-advisor-mcp

📢 Community contribution — This MCP's recommendations are designed for agents in the Gentle AI harness. This is an independent project created to help the community choose the best models for each agent. It is not officially affiliated with Gentle AI or OpenCode.

MCP server that helps LLMs pick the best AI model for each coding agent. Fetches real-time data from the public OpenCode Go/Zen model catalogs and cross-references with OpenRouter benchmarks and reasoning capabilities.

What it does

• Lists all available models in the OpenCode Go and/or Zen subscriptions

• Enriches them with OpenRouter benchmarks (intelligence, coding, agentic scores), pricing, and context window

• Shows reasoning effort support — whether a model supports explicit effort levels (xhigh, high, low) or a simple toggle

• Reads agent selection criteria from the Gentle AI guide so the LLM knows what each agent needs before choosing a model

• Recommends models per agent based on actual data, not guesses

Related MCP server: LLM Benchmark MCP Server

Installation

Option A: npm (recommended)

npm install -g model-advisor-mcp

Then configure OpenCode to use it (see Configuration).

Option B: Manual (development only)

git clone https://github.com/AlonsoSG0/model-advisor-mcp.git
cd model-advisor-mcp
pnpm install
pnpm build

Requirements

• Node.js 18+

No API keys required. The OpenCode (Zen/Go) and OpenRouter model catalogs are public endpoints; the server makes unauthenticated requests and works out of the box.

Configuration

No environment variables or API keys are needed. The server fetches the public model catalogs of OpenCode (Zen and Go) and OpenRouter without credentials.

If you installed via npm (recommended):

{
  "mcp": {
    "model-advisor": {
      "type": "local",
      "command": ["model-advisor-mcp"],
      "enabled": true,
      "timeout": 30000
    }
  }
}

If you are developing locally from a clone, point OpenCode to your local build:

{
  "mcp": {
    "model-advisor": {
      "type": "local",
      "command": [
        "node",
        "/path/to/model-advisor-mcp/dist/server.js"
      ],
      "cwd": "/path/to/model-advisor-mcp",
      "enabled": true,
      "timeout": 30000
    }
  }
}

Quick start and examples

After installing and configuring the MCP:

1. Verify in your terminal that the globally installed executable is available:

   command -v model-advisor-mcp

   The command should return the executable path. If you installed manually, this step does not apply: OpenCode uses the path to dist/server.js configured above.

2. Restart OpenCode so it loads the configuration, then verify the connection status:

   opencode mcp list

   model-advisor should appear connected.

3. Send one of these prompts to your AI agent or orchestrator. They are not terminal commands:

   Using the model-advisor MCP, tell me which models are available in the OpenCode Go and Zen subscriptions.

   Using the model-advisor MCP, give me a low-cost recommendation using only OpenCode Go models.

If the agent can list models or produce a recommendation using MCP data, the connection is working correctly.

Tools

list_available_models

Fetches all AI models from the OpenCode Go and/or Zen catalogs, cross-referenced with OpenRouter.

Parameters:

Returns for each model:

Example reasoning output:

// Model with explicit effort levels
"reasoning": {
  "supportedEfforts": ["xhigh", "high"],
  "defaultEffort": "high",
  "mandatory": false,
  "defaultEnabled": true
}

// Model with toggle only (on/off)
"reasoning": {
  "supportedEfforts": [],
  "defaultEffort": null,
  "mandatory": false,
  "defaultEnabled": true
}

// Model without reasoning
"reasoning": null

get_agent_criteria

Reads agent selection criteria from the Gentle AI guide. Use this before picking a model — each agent has specific needs (context window, reasoning ability, speed, cost).

Parameters:

Agent IDs: gentle-orchestrator, sdd-init, sdd-onboard, sdd-explore, sdd-propose, sdd-spec, sdd-design, sdd-tasks, sdd-apply, sdd-verify, sdd-archive, review-risk, review-readability, review-reliability, review-resilience, review-refuter, jd-judge-a, jd-judge-b, jd-fix-agent

Agent Groups (display recommendations in this order):

1. Orchestrator

2. SDD agents

3. Review (4R)

4. Judgment Day

get_model_benchmarks

Deep-dive into a specific model's OpenRouter data. Useful when list_available_models didn't return benchmarks for a model.

Parameters:

How the LLM uses these tools

The typical workflow:

1. list_available_models → sees what's available, their benchmarks, and reasoning support

2. get_agent_criteria (per agent) → understands what each agent needs

3. get_model_benchmarks (optional) → deeper data on a specific model

4. LLM reasons → matches models to agents based on criteria + benchmarks + cost

Development

# Install dependencies
pnpm install

# Compile TypeScript
pnpm build

# Run directly (for testing)
pnpm start

# Watch mode (auto-reload on changes)
pnpm dev