xyTom/coding-tools-mcp

What it solves

Coding Tools MCP provides a secure, model-neutral runtime for AI agents to interact with a local codebase. It eliminates the need for each AI chat client or agent framework to implement its own file system and shell access tools, providing a standardized set of 18 battle-tested tools for reading, patching, and executing commands within a confined workspace.

How it works

The project implements the Model Context Protocol (MCP), allowing it to act as a server that any MCP-compatible client (such as Claude Desktop, Cursor, or VS Code) can drive. It operates as a workspace-bound runtime that restricts AI agents to a specific directory, preventing unauthorized file access or directory traversal. It supports both stdio and HTTP transports and includes permission modes (safe, trusted, and dangerous) to gate network access, shell expansion, and destructive commands.

Who it’s for

• AI Agent Developers: Those building custom agent loops who want production-grade, secure file and execution tools without hand-rolling them.
• AI Chat Users: Users of tools like Claude Desktop or Gemini CLI who want to turn their chat interface into a coding agent with direct repository access.
• Security-Conscious Developers: Developers who need to run AI agents on untrusted code using the provided Docker sandbox or kernel-level Landlock confinement on Linux.

Highlights

• Model-Neutral: Works with any LLM or MCP client regardless of the vendor.
• Security-First: Features workspace confinement, permission modes, and Landlock kernel-level filesystem confinement on Linux.
• Comprehensive Toolset: Includes tools for file reading/search, structured multi-file patches, Git operations, and interactive command sessions (REPL/debugger).
• Context-Window Optimized: Results are summarized and paginated to reduce token usage and serialized tool-result bytes.
• Flexible Deployment: Can be run locally via uvx or npx, as a remote server via authenticated HTTPS tunnels, or in a disposable Docker sandbox.