ToolPlex Client

This repository contains the official ToolPlex MCP server — the npm package that powers agent interaction with the ToolPlex network. It's the core of ToolPlex Desktop and works with any MCP-compatible client.

ToolPlex is a curated tool ecosystem built for AI agents. With ToolPlex, your agent can:

• Discover 4,000+ high quality, open-source MCP tools (and growing)
• Automatically install and debug complex MCP servers — including fetching READMEs and resolving dependencies
• Call any tool across any installed MCP server without loading every schema into context
• Create and run playbooks (multi-step AI workflows) just by chatting
• Learn from collective agent feedback to improve over time

No complex setup. Just add ToolPlex to your AI client and start automating.

Features

• Agent Tool Discovery — Your agent can search a curated index of MCP servers, filtered by code analysis and popularity signals to find tools that actually work
• Smooth Install Experience — Automatically installs even complex tools, with the agent handling tricky build steps, dependencies, and setup flows behind the scenes
• Seamless Server Activation — Install and uninstall MCP servers at any time without restarting your LLM or resetting the session
• Workflow Memory — Successful tasks are saved as playbooks your agent can search, reuse, and adapt later
• Quality Signals Built-In — Agents report tool usage and failure rates to help down-rank unreliable or broken servers automatically
• Full Agent Control — Use the ToolPlex dashboard to manage server access, shell permissions, file visibility, feedback settings, and more
• Local-First Execution — By default, ToolPlex installs and runs tools on your machine for full speed, privacy, and control

Quick Setup

ToolPlex Desktop (recommended)

Download ToolPlex Desktop — a native app with the ToolPlex client built-in. No configuration needed.

Claude Desktop or other MCP clients

1. Sign up for a ToolPlex AI account and create your first API key.
2. Add this to your MCP client config (e.g. claude_desktop_config.json):

{
  "mcpServers": {
    "toolplex-mcp": {
      "command": "npx",
      "args": ["@toolplex/client"],
      "env": {
        "TOOLPLEX_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}

Works with any AI client that supports MCP.

ToolPlex works best as the only server in your MCP config, since it handles discovery, installation, and management of all other MCP servers on your behalf.

If you must include ToolPlex alongside other servers, be sure to mention toolplex by name when you want to use it. This helps your agent avoid ambiguity when multiple servers are present.

ToolPlex Usage Guide

ToolPlex API users aren't humans — they're AI agents.

You talk to your agent. Your agent talks to ToolPlex. This allows agents to handle the full lifecycle of tool usage — from discovery and installation to configuration, execution, and playbook use — all without human micromanagement.

Ask your AI to activate ToolPlex:

> init toolplex
> start toolplex
> open toolplex

After initializing ToolPlex, just talk to your agent naturally:

Discover Available Tools

> what tools are currently installed?
> show me what ToolPlex servers I can use

Run Installed Tools

> merge these PDFs using the File Merger tool
> fetch the latest bitcoin price from Yahoo Finance
> find some tools to generate my weekly dinner menu
> extract tables from this Excel file using the Excel MCP
> query our BigQuery sales data for Q2 totals

Install New Tools (via search)

> find a tool that can automate browser tests
> is there a server that can visualize data with charts?
> search for a tool that can read and organize research PDFs

Manage Tools

> install the google news server
> uninstall the ffmpeg server
> reinstall the Playwright automation server

Use Playbooks

> run my weekly KPI dashboard playbook
> save this as a playbook

LLM Compatibility

ToolPlex works best with high-context LLMs that support tool-calling:

• Claude Sonnet 4.5, Opus 4.5, Haiku 4.5
• GPT-5, GPT-5 Mini
• Gemini 2.5 Pro, Gemini 3 Pro
• Kimi K2, Kimi K2 Thinking
• Grok 4

Lighter models like DeepSeek V3 or Qwen3 can handle simpler tasks (like running playbooks), but may struggle with complex freeform ToolPlex usage.