A command-line AI chat application that lets you interact with documents using Claude. Built on the MCP (Model Context Protocol) architecture, it connects an AI-powered chat interface to a document server that exposes tools, resources, and prompt templates.

mcpChat implements the complete MCP architecture — a custom MCP server that exposes tools, resources, and prompt templates, paired with a custom MCP client that connects to it over stdio transport. The client feeds retrieved context into a Claude-powered conversational loop, all surfaced through a rich terminal interface.

Built using FastMCP, the server exposes three categories of MCP primitives:

• Tools — read_document and edit_document allow Claude to read and modify documents during a conversation
• Resources — URI-addressable endpoints (docs://documents, docs://documents/{doc_id}) that the client can fetch directly
• Prompts — reusable prompt templates (e.g. format) that inject structured instructions into the message history

A custom MCPClient class manages the full lifecycle of an MCP session:

• Spawns the server as a subprocess and connects via stdio transport
• Initializes and holds a ClientSession for the duration of the app
• Exposes typed methods for list_tools, call_tool, list_prompts, get_prompt, and read_resource

Multiple MCP clients can be registered simultaneously, enabling tool aggregation across servers.

• Document-aware chat — reference any document in your query using @filename and its content is automatically injected as context
• Slash commands — run predefined prompt templates (e.g. /format report.pdf) directly from the CLI
• Tab completion — autocomplete for @documents and /commands powered by prompt_toolkit
• Multi-server support — pass additional MCP server scripts as CLI arguments to extend available tools

The app runs two components side by side:

• MCP Server — manages a document store and exposes read_document and edit_document tools, document resources, and prompt templates
• MCP Client + Claude — the CLI connects to the server over stdio, retrieves context, and sends messages to the Anthropic API

When you type a message, the app resolves any @document references, builds a context-aware prompt, and streams the response from Claude back to the terminal.

1. Create a .env file in the project root:

ANTHROPIC_API_KEY="your-api-key"
CLAUDE_MODEL="claude-sonnet-4-6"

2. Install dependencies and run:

uv venv
source .venv/bin/activate
uv pip install -e .
uv run main.py

Naveen Soni