copx 0.1.3

copx-mcp

✨ CopX - Your Code Project Expert ✨

中文

🚀 Code Project Expert (CopX) is a system designed to answer questions about any codebase.

• Explore, search your codebase agentically (like a human expoert).
• No embedding needed.

Core Features

• Declaration-Based Context: Understands code structure by parsing source files and building a "declaration map" of symbols (functions, classes, variables, etc.). This allows for precise, non-embedding-based retrieval of relevant code context.
• Incremental Updates: Utilizes a hidden Git repository to track file changes and incrementally updates the declaration map, ensuring efficiency for large codebases.
• Hybrid Retrieval (Optional): Can integrate with semantic search (e.g., using LanceDB and embedding models) for queries that benefit from semantic understanding, complementing the primary declaration-based retrieval.
• LLM-Powered Q&A: Leverages Large Language Models (LLMs) to understand questions and formulate answers based on the retrieved context.
• Multi-Language Support: Easily extensible to support various programming languages through Tree-sitter configurations. Currently supports Python, Golang, TypeScript, and JavaScript out-of-the-box.
• Flexible Deployment: Can be run as a FastAPI service or an MCP (Model Context Protocol) server.

Workflow Diagram

Here's a simplified diagram of the agent's workflow:

graph TD
    A[User Question] --> B{CopX Agent};

    subgraph "CopX Core Processing Pipeline"
        B --> D[1 Update/Load Declaration Map];
        D --> E["Symbol Extractor (Tree-sitter) + Git Manager"];
        E --> F[(Declaration Map <br/> .pkl + .git Store)];
        
        D --> G["2 Agentic Code Retrieval <br/> Tools: <br/> - view_file_declaration_map <br/> - view_file_content <br/> - grep_declaration_map <br/> - grep_search <br/> - semantic_search_declaration_map"];
        G -- Uses --> F;
        G -- Optional: Semantic Search --> H[(LanceDB)];
        
        G --> I["3 Formulate Answer (LLM)"];
    end
    
    I --> J[Return Answer];
    J --> A;

Usage Guide

CopX can be run in two modes: as an MCP server or as a FastAPI service.

1. MCP (Model Context Protocol) Server

The MCP server exposes CopX's capabilities as a tool that can be used by MCP-compatible clients (e.g., AI IDEs).

Prerequisites:

• Ensure you have uv installed. You can install it by following the instructions at https://github.com/astral-sh/uv.

To run: Execute the following command in your terminal:

uvx --from copx copx-mcp

This will start the MCP server, registering a tool named "Ask Expert".

Recommended AI IDE Configuration: You can integrate CopX MCP server into your AI IDE (like those supporting MCP) using a configuration similar to this:

{
  "copx": {
    "command": "uvx",
    "args": [
      "--from",
      "copx",
      "copx-mcp"
    ],
    "env": {
      "COPX_MODEL": "ollama/devstral",
      "COPX_DATA_PATH": "~/Documents/Copx",
      "COPX_API_KEY": "YOUR_KEY_HERE",
      "COPX_BASE_URL": "YOUR_LLM_BASE_URL_HERE"
    },
    "disabled": false,
    "autoApprove": [
      "Ask Expert"
    ],
    "timeout": 1800
  }
}

Parameter Explanations:

• COPX_MODEL: Specifies litellm style model name for CopX (e.g., "ollama/devstral", "openai/gpt-4").
• COPX_DATA_PATH: The directory where CopX stores its internal data, such as declaration maps and Git snapshots for projects (e.g., "~/Documents/Copx"). It's recommended to use a persistent path.
• COPX_API_KEY: Your API key for the selected LLM service.
• COPX_BASE_URL: The base URL for your LLM API endpoint.

MCP Tool: Ask Expert

• Description: Get expert answer about project's codebase.
• Arguments:
  • project_path (string, required): The absolute path to the project codebase.
  • question (string, required): The question about the codebase.

2. FastAPI Service

The FastAPI service provides an HTTP endpoint for querying.

Configuration: The query endpoint accepts LLM configuration parameters directly in the request body.

To run: Use Uvicorn (or any ASGI server):

uvicorn copx.copx_fastapi:app --reload --host 0.0.0.0 --port 8000

API Endpoint: POST /query

Request Body (application/json):

{
  "project_path": "/path/to/your/codebase",
  "question": "How does the user authentication work?",
  "model": "openai/gpt-4", // litellm format                
  "base_url": "https://api.example.com/v1", 
  "api_key": "your_llm_api_key",    
  "git_path": "~/.copx_data"        
}

• project_path: Absolute path to the codebase you want to query.
• question: Your question about the codebase.
• model, base_url, api_key: LLM provider details.
• git_path: Directory where CopX will store its cache (declaration maps and git snapshots). It's recommended to use a persistent path like ~/.copx_data or ./.copx_data in your project. If the directory doesn't exist, CopX will create it.

How It Works

CopX processes user queries about a codebase through a multi-stage pipeline:

1. Declaration Map Update/Load:

   • When a query is received for a project, CopX first ensures its declaration map is up-to-date.
   • It uses a ProjectGitManager to track file changes since the last run.
   • A SymbolExtractor (powered by Tree-sitter) parses modified or new files according to language-specific rules.
   • The extracted symbols (functions, classes, variables, with their locations) are stored in the declaration_map. This map is persisted as a .pkl file, and the file versions are tracked in a hidden .git repository within a specified data directory.

2. Code Retrieval:

   • The CodeRetriever node uses the declaration_map to find code segments directly relevant to the user's query based on symbol names and code structure.
   • Optionally, for broader or more abstract queries, it can utilize semantic search (if configured) on a LanceDB index of code (e.g., function bodies).

3. Answer Formulation:

   • The AnswerFormulator node takes the retrieved context (from the declaration map and/or semantic search) and the original question.
   • It then interacts with an LLM to generate a comprehensive answer.

Supported Languages

CopX uses Tree-sitter for code parsing and supports languages via configuration files found in src/copx/symbal_extractor/configs/. Currently supported:

• Python (.py)
• Golang (.go)
• JavaScript (.js, .jsx)
• TypeScript (.ts, .tsx)

Adding Support for a New Language

To add support for a new programming language:

1. Install Tree-sitter grammar: Find the Python binding for the Tree-sitter grammar (e.g., tree-sitter-ruby) and install it:

   pip install tree-sitter-newlanguage
   

2. Create a configuration file: In the src/copx/symbal_extractor/configs/ directory, create a new JSON file (e.g., newlanguage.json). This file needs to specify:

   • tree_sitter_module: The name of the installed Python module for the language (e.g., tree_sitter_newlanguage).
   • language_accessor_name (optional): If the language object in the module is not accessed by module.language, specify the correct accessor (e.g., get_language).
   • extraction_rules: An array of rules defining which AST node types correspond to which symbols (function, class, variable, etc.) and how to extract their names. Refer to existing config files (e.g., py.json, go.json) for examples.

   Example newlanguage.json:

   {
     "tree_sitter_module": "tree_sitter_newlanguage",
     "extraction_rules": [
       {
         "node_type": "function_definition",
         "name_field": "name",
         "symbol_type": "Function"
       },
       {
         "node_type": "class_definition",
         "name_field": "name",
         "symbol_type": "Class"
       }
       // ... more rules
     ]
   }
   

Installation

1. Clone the repository:

   git clone <repository_url>
   cd copx
   

2. Install dependencies: It's recommended to use a virtual environment.

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   pip install -r requirements.txt # Or: uv pip install -r requirements.txt
   

   You'll need to install Tree-sitter language bindings for the languages you intend to support, for example:

   pip install tree-sitter-python tree-sitter-go tree-sitter-javascript tree-sitter-typescript