Skip to content

use-mcp-servers.md

Latest commit

 

History

History
178 lines (118 loc) · 5.18 KB

use-mcp-servers.md

File metadata and controls

178 lines (118 loc) · 5.18 KB

How to Use MCP Servers

This guide shows you how to set up and interact with Model Context Protocol (MCP) servers to extend gpt2099's capabilities.

Overview

MCP servers are external tools that provide additional capabilities like file editing, web search, or database access. gpt2099 integrates with these servers through the cross.stream generator pattern, allowing you to experiment with and understand server capabilities before using them in conversations.

Setting Up MCP Servers

1. Register an MCP Server

gpt mcp register filesystem "npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/directory"

This spawns the server as a cross.stream generator, making its tools available for use.

2. Discover Available Tools

gpt mcp tool list filesystem

This shows you what tools the server provides:

──#──┬───────────name────────────┬─────────────────────────────────────
 0   │ read_file                 │ Read the contents of a file
 1   │ read_multiple_files       │ Read multiple files at once
 2   │ write_file                │ Write content to a file
 3   │ create_directory          │ Create a new directory
 ...

3. Test Tools Directly

Before using tools in conversations, test them manually:

gpt mcp tool call filesystem read_file {path: "/path/to/file.txt"}

Test tools to understand their input format and expected output.

Using MCP Servers in Conversations

Basic Usage

Once registered, specify servers when making requests:

"Read the contents of config.json and explain its structure" | gpt --servers [filesystem] -p milli

The LLM will automatically use the appropriate tools from the filesystem server.

Multiple Servers

You can use multiple servers simultaneously:

gpt mcp register web-search "npx -y @modelcontextprotocol/server-brave-search"

"Research current best practices and update our README.md file" | gpt --servers [web-search, filesystem] -p kilo

Interactive Tool Execution

When the LLM wants to use a tool, you see the proposed tool call and have several options:

┌─────────────┬──────────────────────────────────────────────────────────┐
│ name        │ filesystem___write_file                                  │
│ input       │ {path: "config.json", content: "{\n  \"updated\": true}"}│
└─────────────┴──────────────────────────────────────────────────────────┘

Execute?
> yes
  no: do something different
  no
  activate: yolo

Options:

  • yes - Execute the tool call as proposed
  • no: do something different - Provide custom input or alternative response
  • no - Skip this tool call and stop
  • activate: yolo - Enable YOLO mode for automatic execution

YOLO Mode

YOLO mode automatically executes tool calls without prompting:

# Enable YOLO mode via environment variable
$env.GPT2099_YOLO = true
"Update all config files" | gpt --servers [filesystem] -p milli

Or activate it during a conversation by selecting "activate: yolo" when prompted. Once activated, all subsequent tool calls in the session execute automatically.

Custom Tool Responses

When you select "no: do something different", you can provide custom input:

Enter alternative response: The file already exists and shouldn't be modified

This sends your custom response as the tool result, allowing you to guide the conversation without executing the actual tool.

Server Management

List Active Servers

gpt mcp list

Shows all currently running MCP servers.

Server Lifecycle

Servers run as cross.stream generators and will:

  • Automatically restart if they crash
  • Be available across multiple conversations
  • Terminate when the cross.stream session ends

Common MCP Servers

Here are some useful MCP servers to try:

File System Operations:

gpt mcp register filesystem "npx -y @modelcontextprotocol/server-filesystem /workspace"

Web Search:

gpt mcp register brave-search "npx -y @modelcontextprotocol/server-brave-search"

Git Operations:

gpt mcp register git "npx -y @modelcontextprotocol/server-git"

Database Access:

gpt mcp register postgres "npx -y @modelcontextprotocol/server-postgres postgresql://user:pass@localhost/db"

Experimentation Workflow

The recommended approach for working with new MCP servers:

  1. Register the server
  2. List available tools to understand capabilities
  3. Test individual tools with gpt mcp tool call
  4. Use in conversations with --servers flag
  5. Iterate based on results

Hands-on experimentation builds understanding of server capabilities and effective prompts.

See the commands reference for complete MCP command options.