Llm Advisor MCP

What Daichi-Kudo/llm-advisor-mcp MCP server does

Daichi-Kudo/llm-advisor-mcp MCP server helps an AI assistant evaluate available LLM and VLM models using current public data. It brings together pricing, context limits, modalities, release dates, benchmark scores, and capability indicators in a single MCP interface. The results are formatted as short Markdown tables or structured sections instead of large raw-data responses, which keeps tool output relatively small.

The server supports four decision-making workflows. Developers can inspect one model in detail, rank models by a selected category, compare two to five models side by side, or request three recommendations based on a use case and constraints. Supported ranking categories include coding, math, vision, general quality, cost effectiveness, open source, speed, context window, and reasoning.

How it works

The server reads data from OpenRouter, SWE-bench, LM Arena, OpenCompass VLM, and Aider Polyglot. These sources provide model pricing and metadata, coding results, human-preference ratings, vision benchmarks, and multilingual coding scores. Data is retrieved from public APIs or published leaderboards without authentication.

Source data is cached for different periods. OpenRouter data has a one-hour cache lifetime, while the benchmark sources generally use six-hour cache periods. Recommendations combine benchmark weighting, pricing, capability matches, and a freshness adjustment for recently released models. They should therefore be treated as data-driven guidance based on the available sources, not as a substitute for testing a model in a specific application.

Setup and configuration

Daichi-Kudo/llm-advisor-mcp runs locally over the MCP stdio transport. Node.js 18 or newer is required. Install and launch it with:

npx -y llm-advisor-mcp

For Claude Desktop, Cursor, or Windsurf, add a server entry whose command is npx and whose arguments are -y and llm-advisor-mcp. Claude Code can also add the same server through its MCP command. No API keys, registration, .env file, or other credential configuration is needed.

Tools and capabilities

Daichi-Kudo/llm-advisor-mcp exposes these tools:

• get_model_info accepts a model ID or partial name and returns pricing, benchmarks, percentile ranks, context and capability details, plus an optional API example. Examples can use the OpenAI SDK, cURL, or Python requests.
• list_top_models ranks models for a selected category. Optional filters include result count, minimum context length, and minimum release date.
• compare_models compares two to five model IDs across pricing, context, maximum output, benchmarks, capabilities, open-source status, and release date. Best values are highlighted in the result.
• recommend_model returns three ranked choices for coding, math, general, vision, creative, reasoning, or cost-effective workloads. Filters can require price limits, context size, vision, tools, open-source status, or a release date.

Responses are intentionally compact: the README describes typical outputs of roughly 250 to 400 tokens depending on the tool.

Limitations and notes

The server reports only the models and measurements represented by its five data sources. Coverage and freshness vary by source, and benchmark scores may not predict performance for every workload. Recommendation scoring uses fixed factors described by the project, including weighted benchmarks, price, capability bonuses, and release freshness.

The API code example returned by get_model_info is a usage template for the selected provider route; using that example may require a separate provider credential even though the MCP server itself does not require one. The project is distributed under the MIT license and requires Node.js 18 or newer.