Persistent-Code MCP Server with LlamaIndex

A Model Context Protocol (MCP) server that creates and maintains a semantic knowledge graph of code generated by Claude. Powered by LlamaIndex, this allows maintaining context across sessions with advanced semantic search capabilities without requiring the entire codebase to be present in the context window.

Problem & Solution

When developing software with Claude:

Context windows are limited, making it difficult to work with large codebases
Previous code context is lost between sessions
Claude lacks persistent understanding of project structure
Redundant explanation of code is required in each session
Maintaining implementation consistency is challenging

Persistent-Code solves these problems by:

Creating a knowledge graph of code components and their relationships
Tracking implementation status of each component
Providing tools to navigate, query, and understand the codebase
Assembling minimal necessary context for specific coding tasks
Maintaining persistent knowledge across chat sessions

LlamaIndex Integration

Persistent-Code leverages LlamaIndex to provide enhanced semantic understanding:

Semantic Search: Find code components based on meaning, not just keywords
Vector Embeddings: Code is embedded into vector space for similarity matching
Knowledge Graph: Relationships between components are tracked semantically
Contextual Retrieval: Related code is retrieved based on semantic relevance

This integration allows Claude to understand your codebase at a deeper level:

Find functions based on what they do, not just what they're called
Get more relevant code components when preparing context
Better understand the relationships between components
More accurately retrieve examples of similar implementations

Installation

Prerequisites

Python 3.10 or higher
UV package manager (recommended) or pip

Setting Up

# Clone repository
git clone https://github.com/your-username/persistent-code-mcp.git
cd persistent-code-mcp

# Set up environment with UV
uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
uv pip install -r requirements.txt

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

Usage

Initializing a Project

python -m persistent_code init --project-name "YourProject"

Starting the Server

python -m persistent_code serve --project-name "YourProject"

Configuring Claude for Desktop

Edit your Claude for Desktop config file:
- Location: ~/Library/Application Support/Claude/claude_desktop_config.json
- Add the following configuration:

{
  "mcpServers": {
    "persistent-code": {
      "command": "path to python in venv",
      "args": [
        "-m",
        "persistent_code",
        "serve",
        "--project-name",
        "default"
      ],
      "cwd": "persistent-code-mcp",
      "env": {
        "PYTHONPATH": "abs path to persistent-code-mcp"
      }
    }
  }
}

Restart Claude for Desktop
Connect to your MCP server by asking Claude about your code

Available Tools

Knowledge Graph Management

add_component: Add a new code component to the graph
update_component: Update an existing component
add_relationship: Create a relationship between components

get_component: Retrieve a component by ID or name
find_related_components: Find components related to a given component
search_code: Search the codebase semantically

Status Management

update_status: Update implementation status of a component
get_project_status: Retrieve implementation status across the project
find_next_tasks: Suggest logical next components to implement

Context Assembly

prepare_context: Assemble minimal context for a specific task
continue_implementation: Provide context to continue implementing a component
get_implementation_plan: Generate a plan for implementing pending components

Code Analysis

analyze_code: Analyze code and update the knowledge graph

Example Workflow

Initialize a project:
python -m persistent_code init --project-name "TodoApp"
Start the server:
python -m persistent_code serve --project-name "TodoApp"
Ask Claude to design your project:
Can you help me design a Todo app with Python and FastAPI? Let's start with the core data models.
Claude will create components and track them in the knowledge graph
Continue development in a later session:
Let's continue working on the Todo app. What's our implementation status?
Claude will retrieve the current status and suggest next steps
Implement specific components:
Let's implement the task completion endpoint for our Todo app
Claude will retrieve relevant context and provide consistent implementation

Using Semantic Search

With the LlamaIndex integration, you can now use more natural language to find components:

Find me all code related to handling task completion

Claude will use semantic search to find relevant components, even if they don't explicitly contain the words "task completion".

Running the LlamaIndex Demo

We've included a demo script to showcase the semantic capabilities:

# Activate your virtual environment
source .venv/bin/activate  # or source venv/bin/activate

# Run the demo
python examples/llama_index_demo.py

This will demonstrate analyzing a Calendar application and performing semantic searches for functionality.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Persistent-Code MCP Server

Integrations