mcp-client-cli Documentation

Run prompts directly via the llm command. You can pass a string as an argument or pipe content from other commands or files.

Warning: If your prompt contains shell special characters like &, |, or ;, wrap the prompt in quotes to prevent the shell from interpreting them.

# Direct prompt
llm "What is the capital city of North Sumatra?"

# Piping from echo
echo "What is the capital city of North Sumatra?" | llm

# Piping from a file
cat instructions.txt | llm "West Java"

Access predefined prompt templates using the p prefix followed by the template name and optional arguments.

• Use llm --list-prompts to see all available templates.
• Syntax: llm p <template_name> [args]

# List templates
llm --list-prompts

# Use templates
llm p review
llm p commit
llm p yt url=https://youtube.com/...

Install the mcp-client-cli package using pip to run LLM prompts and implement Model Context Protocol (MCP) clients from your terminal.

pip install mcp-client-cli

The cb command allows you to interact with text or images currently in your system clipboard.

• llm cb: Processes the clipboard content directly.
• llm cb "<question>": Analyzes the clipboard content using the provided question.
• llm cb c "<question>": Combines clipboard usage with conversation continuation.

Platform Requirements:

• Windows: Uses PowerShell (built-in).
• macOS: Uses pbpaste (built-in) for text; pngpaste (optional, brew install pngpaste) for images.
• Linux: Requires xclip (sudo apt install xclip).
• WSL: Accesses Windows clipboard via powershell.exe.

# Process clipboard text
llm cb

# Ask a question about clipboard content
llm cb "What language is this code written in?"

# Continue conversation with clipboard content
llm cb c "Tell me more about what you see"

When an LLM decides to use an MCP tool, the CLI will display the tool call details.

• Manual Confirmation: If a tool is listed in the requires_confirmation section of your config, you will be prompted with Confirm tool call? [y/n]:.
• Bypass Confirmation: Use the --no-confirmations flag to automatically approve all tool calls.
• Scripting Mode: Use the --no-intermediates flag in bash scripts to suppress intermediate tool/AI messages and only output the final response.

# Standard tool call (may prompt for confirmation)
llm "What is the top article on hackernews today?"

# Bypass all confirmations
llm --no-confirmations "What is the top article on hackernews today?"

# Use in scripts (suppresses intermediate output)
llm --no-intermediates "What is the time in Tokyo right now?"

The MCP Client CLI looks for its configuration in one of two locations:

1. ~/.llm/config.json (in your user's home directory for global settings).
2. mcp-server-config.json (in your current working directory for project-specific settings).

The file uses JSON format and supports comments using the // syntax.

To continue the context of your last interaction, use the c prefix (note the space) before your message.

$ llm "Hello!"
[AI Response]
$ llm c "What did I just say?"

The McpTool class is a specialized BaseTool implementation that wraps an MCP tool. When called via its asynchronous _arun method, it executes the tool on the remote MCP server using the active ClientSession.

Key Behaviors:

• Automatic Session Management: If the session is not active when the tool is called, it attempts to start the session via the parent toolkit.
• Error Handling: If the MCP server returns an error (result.isError), the tool raises a ToolException containing the error content.
• Output: Returns the tool's result as a JSON-decoded string.

The llm object allows you to define the provider and model parameters. If api_key is omitted, the client will attempt to use the LLM_API_KEY or OPENAI_API_KEY environment variables.

"llm": {
  "provider": "openai",
  "model": "gpt-4o-mini",
  "api_key": "your-api-key-here",
  "temperature": 0
}

Prompt templates allow you to reuse structured prompts. You can list available templates using llm --list-prompts. To use a template, use the p prefix followed by the template name and any required arguments.

If a template contains placeholders like {arg}, you must provide the values in the order they appear in the template.

# If 'review' template is: 'Review this: {text}'
llm p review "my code snippet"