Inferbench

Overview

Inferbench is an MCP server that measures local LLM inference performance by benchmarking tokens per second on your own hardware. It integrates directly into AI assistants as MCP tools, letting you test inference speed across different models and configurations without leaving your editor or chat interface.

The server runs locally via the inferbench-cli package using uvx, communicating over stdio with no authentication needed. It ships at v0.1.7 under the Apache-2.0 license with source code available at RudrenduPaul/InferBench on GitHub.

Inferbench appears in the AI & Agents category and works with MCP clients including Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Zed, and others. If you maintain this project, you can claim the listing to verify ownership and keep information up to date.

Installation

// claude_desktop_config.json
{
  "mcpServers": {
    "inferbench": {
      "command": "uvx",
      "args": [
        "inferbench-cli"
      ]
    }
  }
}