llm-mcp-plugin

LLM MCP Plugin

IMPORTANT: This is a very early version of the MCP wrapper for llm and Bespoken, largely written by Claude and barely tested. Proceed with caution.

A plugin for the LLM command-line tool that enables using Model Context Protocol (MCP) servers as toolboxes in conversations.

Features

• Dynamic MCP Server Integration: Connect to any MCP server and expose its tools as LLM toolbox methods
• Multiple Transport Support: Works with stdio, SSE, and HTTP-based MCP servers
• Automatic Tool Discovery: Automatically converts MCP tools into callable methods
• Tool Filtering: Filter exposed tools by name using include/exclude lists for enhanced security
• Resource & Prompt Access: Expose MCP resources and prompts through the toolbox
• Configuration-Driven: Easy setup for different MCP servers via configuration files
• STDERR Handling: Control how MCP server error output is handled (disable, redirect to file, or display in terminal)

Installation

uv add git+https://github.com/d-lowl/llm-mcp-plugin.git

Development

Setup

To set up the development environment:

# Install all dependencies (including linting tools and pre-commit)
uv sync

# Install pre-commit hooks for automatic code quality checks
uv run pre-commit install

# (Optional) Run pre-commit against all files
uv run pre-commit run --all-files

The pre-commit configuration includes:

• Black - Code formatting
• isort - Import sorting
• mypy - Type checking
• flake8 - Linting
• bandit - Security checks

Quick Start

Usage in Bespoken:

mcp_toolbox = MCPToolbox(
    config=MCPServerConfig(
        name="notion",
        transport="stdio",
        command="npx",
        args=["-y", "@notionhq/notion-mcp-server"],
        env={
            "OPENAPI_MCP_HEADERS": f'{{"Authorization": "Bearer {NOTION_API_KEY}", "Notion-Version": "2022-06-28" }}'
        }
    )
)

chat(
    model_name="anthropic/claude-3-5-sonnet-20240620",
    tools=[TodoTools(), mcp_toolbox],
    system_prompt=SYSTEM_PROMPT,
    debug=True,
    slash_commands={
        "/thinking": "Let me think through this step by step:",
        "/role": set_role,
        "/whoami": whoami,
    },
)

(I haven't done any testing in bare llm yet, but it should work.)

Examples

This repo contains an example Bespoken agent script for Notion.

export NOTION_API_KEY=your-notion-api-key
export ANTHROPIC_API_KEY=your-anthropic-api-key

./notion_example.py

Tool Filtering

The plugin supports filtering tools by name to enhance security and control which tools are exposed. You can use:

• Include filters: Only expose specific tools (whitelist approach)
• Exclude filters: Hide specific tools (blacklist approach)
• Combined filters: Use both include and exclude lists together

CLI Examples

# Only expose read and write operations from a filesystem server
llm mcp add filesystem \
  --transport stdio \
  --command python \
  --args -m \
  --args mcp.server.filesystem \
  --args /safe/directory \
  --tool-include read_file \
  --tool-include write_file \
  --tool-exclude delete_file

# Exclude dangerous administrative tools
llm mcp add api-server \
  --transport http \
  --url http://localhost:8000/mcp \
  --tool-exclude admin_reset \
  --tool-exclude clear_database \
  --tool-exclude delete_all

Configuration File

Tool filters are also saved in the configuration file:

{
  "servers": {
    "safe-filesystem": {
      "transport": "stdio",
      "command": "python",
      "args": ["-m", "mcp.server.filesystem", "/safe/dir"],
      "tool_filter_include": ["read_file", "write_file"],
      "tool_filter_exclude": ["delete_file"]
    }
  }
}

Supported MCP Servers

This plugin works with any MCP server that follows the Model Context Protocol specification. Some examples:

• Toggl Track MCP - Time tracking integration
• Filesystem MCP - File system operations
• Git MCP - Git repository management
• SQLite MCP - Database operations

License

MIT License. See LICENSE for details.

Contributing

Contributions welcome! Please read CONTRIBUTING.md for guidelines.