plandex

Plandex is a terminal-based AI coding agent designed for large coding tasks and real-world projects — the kind that span many steps and touch dozens of files. Rather than just suggesting a snippet, it aims to plan and execute substantial pieces of work end-to-end, while keeping you in control.

The way it works is built around a sandbox review model. AI-generated changes are kept separate from your actual project files until you decide to apply them, so you can review every modification, roll things back, and debug without leaving a mess behind. Command execution is gated in a similar way. It uses tree-sitter (a code-parsing tool) to build project maps, which lets it stay reliable across large codebases — the README claims an effective context window of around 2 million tokens with its default model pack, and supports indexing directories of 20 million tokens or more.

A key idea is that you can mix and combine models from providers like Anthropic, OpenAI, Google, and open source options, using curated "model packs" that trade off capability, cost, and speed. Other features include automated debugging of terminal commands and browser apps, full version control with branches for exploring alternatives, and Git integration with commit-message generation.

You would reach for Plandex when other AI coding tools break down on bigger projects — large files, complex refactors, or features touching many files at once. You can run it in fully autonomous mode where it loads files, plans, implements, runs commands, and debugs on its own, or use step-by-step review. It is written in Go, installs with a one-line shell command, and supports Windows only through WSL.