onetool-mcp 1.0.0b1

One MCP, unlimited tools.

One tool to rule them all, one tool to find them, one tool to bring them all, and in the development bind them.

v1.0.0 Pre-Release - API stable, actively tested.

OneTool is a local-first MCP server that exposes a single run tool for code execution, giving your AI assistant access to unlimited capabilities through one interface.

The Problem

Connect 5 MCP servers and you've burned 55K tokens before the conversation starts. Connect 10+ and you're at 100K tokens. Your AI gets worse as you add more tools - that's not a bug, it's how context windows work.

The Solution

98.7% fewer tokens. Same accuracy. 10x lower cost.

Instead of loading 50 separate tool schemas, you write Python directly:

__ot brave.search(query="AI trends 2026")

No JSON schema parsing. No tool selection loops. No hoping the model guesses correctly. You write explicit code to call APIs - deterministic, visible, no hidden magic.

Based on Anthropic's research, which found token usage dropped from 150,000 to 2,000 when presenting tools as code APIs.

Core Capabilities

• 30-second setup - Install with uv or pip
• Drop-in extensibility - Add a Python file, get a new pack
• AST security - All code validated before execution
• Benchmark harness - Test LLM + MCP combinations with bench

Batteries Included with 100+ Tools

See Tool Reference for the complete list of packs and tools.

Installation

uv tool install onetool-mcp

Or with pip: pip install onetool-mcp

With optional dependencies (for convert, excel, code search):

uv tool install onetool-mcp \
  --with pymupdf --with python-docx --with python-pptx \
  --with openpyxl --with Pillow --with duckdb --with openai

Add to Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "onetool": {
      "command": "onetool"
    }
  }
}

Extending

Drop a Python file, get a pack. No registration, no config:

# tools/mytool.py
pack = "mytool"

def search(*, query: str) -> str:
    """Search for something."""
    return f"Results for: {query}"

Why this approach

LLMs write Python instead of parsing JSON schemas. You see what's being called. 2K tokens instead of 150K. Adding your own packs is just dropping in a file.

Documentation

• Why OneTool - The problem and our solution
• Getting Started - 2-minute setup
• Tools Reference - All built-in tools
• Extending - Create your own tools

References

• Code Execution with MCP - Anthropic Engineering
• Context Rot - Chroma Research

Licensing

GPLv3 - Will transition to MIT at v2.0. Contribute via PRs to help us get there.

Support

If you use or like this project, please consider buying me a coffee: