Repomix

An open-source tool that packages the entire codebase into an AI-friendly format, supports local and remote repositories, can output XML, Markdown, JSON or plain text, and has built-in token counting, compression and security filtering.

AI Developer Tools#Open Source#Codebase#Packaging

Updated: 2026-10-26

What is Repomix?

Repomix solves a very specific problem: fitting an entire codebase into an AI's context window. It packages local or remote repositories into a single file, with optional output formats including XML, Markdown, JSON, and plain text. Specifically designed for models like Claude, ChatGPT, and Gemini, it also integrates seamlessly with MCP and code review workflows and has been nominated for the Open Source Awards.

Who is it for?

Developers who need to provide project context to AI for code review and refactoring suggestions; teams that want to generate Knowledge Datapacks for LLMs; and people who use LLM-powered code generation workflows and require stable context input.

Key features

Multi-format Output

XML, Markdown, JSON, and plain text are optional to adapt to the input preferences of different models; parsable formats can also be output for program consumption.

AI Optimization

By default includes file summaries and directory structures, with an optional line number display feature. This allows models to first build a global understanding before delving into details, thereby improving the quality of comprehension.

Git Awareness

Respects .gitignore by default, only packaging the truly necessary source code; include/ignore rules use glob patterns to precisely control which files are included in the package, avoiding clutter. However, custom include rules may bypass .gitignore, so pay attention to boundaries during configuration.

Token Counting & Compression

Estimates the size before packaging to avoid exceeding the model's context limit; provides compression options (remove comments, delete blank lines) to slim down large repositories, but compression reduces code readability and should be used with caution in code review scenarios.

Security Filtering

Designed for sharing and submission scenarios, it helps identify sensitive information (Security-Focused) and reduces the risk of exposing keys in conversations.

CLI and Web Version

The command-line interface can be integrated into daily workflows and scripts (the official website provides a Quick Start and a Power Users Guide), and there is also an online version where you can directly paste the repository address to use it. You can even adjust the format and filtering options directly on the webpage before copying the results.

How to use Repomix

When conducting code reviews, packaging in batches by module is more stable than dumping the entire repository at once.
first define the include/ignore rules to ensure consistent boundaries for each packaging.
When a repository needs to be reused for a long time, codify the packaging commands into scripts to guarantee uniform output format across the team.
The official team has also summarized two practical use cases: the LLM code generation workflow and Knowledge Datapacks.
By matching your own scenarios, you can quickly find the appropriate configuration.
Sending a packaged project context to new team members can also significantly speed up their onboarding and understanding.

Things to know

Even large repositories may still bloat beyond the context limit even after compression, so be sure to check the token count before feeding them to the model.

It packages a snapshot of the code, and the model cannot see the real-time state of the repository; you will need to repackage it after any changes.

Open-source code is safe to package and share directly, while proprietary code requires caution to avoid exposing sensitive information in public settings.

Packing an ultra-large repository takes a long time, so it is recommended to process it in batches by module; the web version may have restrictions on ultra-large repositories, and for giant repositories, it is advisable to switch to local packaging via the CLI.

The output is code snapshot-style corpus; the model's understanding relies on the comments and naming quality of the code itself, and projects with chaotic naming will see reduced effectiveness.

Binary resources such as images, audio and video files will not be included in the text output. Contexts involving such files need to be handled separately; the output is mainly in plain text, and format information such as code highlighting is limited.