rusiaaman/wcgw

What it solves

It allows AI agents (like Claude) to interact directly with a user's local machine to write, build, and run code. It bridges the gap between a chat interface and a local development environment by providing a secure-ish but powerful way for an LLM to execute shell commands and edit files without needing a manual copy-paste workflow.

How it works

wcgw is an MCP (Model Context Protocol) server that exposes a set of tools to an LLM. These tools allow the agent to execute bash commands, read and edit files using search-and-replace blocks, and manage project context. It supports interactive shell sessions (including sending keystrokes) and can be integrated with VS Code via an extension to allow users to attach to the same terminal the AI is using.

Who it’s for

Developers who want an agentic coding experience where the AI can autonomously iterate on code, run compilers, and fix errors on their local machine.

Highlights

• Interactive Shell: Supports arrow keys, interrupts, and ANSI escape sequences, allowing the AI to handle interactive commands.
• Smart File Editing: Uses Aider-like search-and-replace for efficiency and supports incremental edits for large files to avoid token limits.
• Syntax Validation: Automatically reports syntax errors back to the LLM after a file edit so the agent can correct its mistakes.
• Operational Modes: Includes 'Architect' (read-only planning), 'Code-writer' (restricted editing), and 'wcgw' (full access) modes.
• Context Checkpointing: The ContextSave tool allows the agent to save the current state and relevant files into a single file for later resumption or knowledge transfer.
• Terminal Multiplexing: Integrates with screen and a VS Code extension, letting users watch or interact with the AI's terminal in real-time.