Repomix

Repomix packs a whole repository into one AI-friendly file, and its MCP mode lets your agent do that on demand, then grep and read the packed output in slices. It's the simplest way to hand an agent an entire codebase, yours or any GitHub repo, with no index or database to maintain. Reach for it for whole-repo reviews and onboarding; for day-to-day symbol-level edits, a language-server tool fits better.

Highlights

• CLI and MCP server in one — the repomix CLI writes a packed file you can hand to any model, and repomix --mcp exposes the same packing to an agent over stdio.
• Local or remote — pack_codebase packs a directory, and pack_remote_repository accepts a GitHub URL or user/repo.
• Read in slices — grep_repomix_output and read_repomix_output (with line ranges) let the agent search a pack instead of loading all of it, and attach_packed_output attaches an existing pack.
• Tree-sitter compression — optional; Repomix says it "reduces token usage by ~70% while preserving semantic meaning."
• Secret-aware — Repomix refuses content that matches secret formats (a Secretlint-based heuristic), and --sandbox mode confines the file tools to one workspace root.
• Built for Claude Code — official plugins (repomix-mcp, repomix-commands, repomix-explorer), and the CLI can generate Claude Agent Skills into .claude/skills/<name>/ with --skill-generate.

In an AI-assisted workflow

Add the MCP server to Claude Code with the command from the README:

claude mcp add repomix -- npx -y repomix --mcp

Or install the official plugin instead:

/plugin marketplace add yamadashy/repomix
/plugin install repomix-mcp@repomix

Then ask for packs by task:

Pack this repository with Repomix, then grep the output for every place we read environment variables.
Pack the remote repo for the HTTP client we depend on and explain how its retry logic works.

If you haven't installed from a plugin marketplace before, our guide to Claude Code plugins walks through how it works.

How it compares

Serena is the main alternative for code context, with a different shape: Repomix takes a whole-repo snapshot and searches it with regex grep, while Serena makes incremental, symbol-level queries and edits through language servers, and Repomix has no editing tools. Claude Context sits between them, querying a persistent vector index instead of a snapshot, at the cost of an embedding provider and a database. Repomix isn't a docs index like Context7: pack_remote_repository can pull a library's source, but you get raw code, not curated, version-matched docs.

Good to know

Repomix is MIT licensed, maintained by Kazuki Yamada, and very active: v1.18.0 shipped in August 2026, commits continued into September 2026, and the repo has about 28k GitHub stars as of September 2026. It requires Node.js 22 or newer (its npm engines field says >=22.0.0), or Docker via the ghcr.io/yamadashy/repomix image. It runs locally over stdio, with no hosted endpoint, account, or API key.

The source also registers a generate_skill MCP tool that the README's tool list leaves out; sandbox mode disables skill generation along with remote packing. We couldn't confirm which languages Tree-sitter compression covers, so check it on your own stack before counting on the savings. For the rest of a context stack, see the best MCP servers in 2026, and for scopes and .mcp.json, adding MCP servers to Claude Code.