MCPunk

MCPunk is an MCP server that enables developers to explore and understand local codebases through direct conversation without using vector embeddings. It connects an MCP-compatible client, such as Claude Desktop, to local project directories and source code files. By breaking files into logical chunks—such as individual Python functions, classes, imports, Vue blocks, or Markdown sections—MCPunk allows language models to search and query code hierarchically. Software engineers and code reviewers use MCPunk to navigate unfamiliar repositories, inspect pull requests, trace dependencies across multiple local projects, and inspect recent git changes. Instead of relying on opaque vector search or semantic embeddings, MCPunk relies on transparent text searches and structural chunking. The language model navigates projects step-by-step by identifying candidate files, locating relevant chunk identifiers, and retrieving full chunk contents as needed. This approach, referred to as roaming RAG, keeps context windows efficient while making search trails explicit and auditable to the user during debugging or technical code investigations.

Category: Developer Tools & Code Intelligence

Tags: code analysis, codebase, search, source-code

Visit MCPunk

How to install and configure MCPunk

1. Install uv on your system following Astral's installation instructions. 2. Open your claude_desktop_config.json configuration file. 3. Add the MCPunk server definition under the mcpServers configuration block: json { "mcpServers": { "MCPunk": { "command": "uvx", "args": ["mcpunk"] } } } Note: If uvx is not detected in your environment path, replace uvx with its absolute executable path. 4. Restart Claude Desktop to load the new server tools. 5. Ask your model to configure a repository path using configure_project to begin searching chunks.

What you can do with MCPunk

• Inspecting unfamiliar codebases by having the model index local files and retrieve relevant function or class chunks using structural text searches. - Performing pull request reviews by running git diff comparisons against target reference branches and checking modified file chunks. - Tracing cross-project dependencies by configuring multiple local repositories in the same chat session to query related functions and imports. - Investigating recent branch activity by parsing local git reflog outputs to list recently checked-out branches before inspecting code differences.

Key facts

• https://github.com/jurasofish/mcpunk
• Developer Tools & Code Intelligence, Files, Documents & PDFs
• code analysis, codebase, search, source-code

Part of MCP Servers

How do I install MCPunk?

To install MCPunk, ensure you have uv installed on your system. Next, open your Claude Desktop configuration file and add an entry for MCPunk under mcpServers, specifying uvx as the command and mcpunk as the argument. Save the configuration and restart Claude Desktop. Once loaded, the server exposes tools to configure local projects and search code chunks directly in chat.

What can MCPunk do?

MCPunk splits local repository files into structural chunks, such as Python classes, functions, Vue components, and Markdown sections. It provides tools for language models to list project files, perform exact text searches across chunks, fetch specific code contents, and inspect git diffs. This enables conversational codebase exploration and code review without building vector embeddings.

Which MCP clients work with MCPunk?

MCPunk works with Claude Desktop and any standard client implementing the Model Context Protocol. Because it runs over standard MCP interfaces using command-line runners like uvx, developers can register the server in any compatible client application that supports external MCP tool configuration.

How does MCPunk search code without embeddings?

MCPunk employs a roaming RAG pattern instead of vector databases. When a project is configured, files are segmented into logical chunks such as functions or sections. The LLM executes exact text search queries across chunk contents, reviews matching chunk identifiers, and requests full content only for relevant pieces, preserving context while showing the exact search trail.