diffthing

Your agent writes code. You still own judgment.

Local-first diff review for AI-assisted development. diffthing turns working-tree changes into a prioritized walkthrough, keeps review state stable while agents keep editing, and stages files only after human approval.

AI organizes and executes. It never approves code.

Usage

Run inside any Git repository:

npx diffthing

  diffthing 0.1.0
  reviewing /path/to/project against HEAD
  llm       your-llm (your login)
  ✓ ready   0 files, 0 changes, 1 AI-organized scopes

  open  https://local.diffthing.dev:58826/#port=58826&token=…

Open the printed URL to review uncommitted changes against HEAD. local.diffthing.dev resolves to 127.0.0.1 — the review UI runs entirely on your machine, served over HTTPS from the embedded local daemon. Nothing to install or trust.

Install globally instead:

npm install -g diffthing
diffthing

Options

diffthing [OPTIONS]

--base <BASE>  Diff base; default HEAD
--offline      Serve over plain HTTP on 127.0.0.1 instead of HTTPS via
               local.diffthing.dev
--port <PORT>  Fixed port; default first free port
--repo <REPO>  Repository root; default current directory
--llm <LLM>    claude | codex | gemini | kimi | qwen | opencode | none | auto

The npm build ships a prebuilt binary with the review UI embedded and serves it over HTTPS via local.diffthing.dev by default. If your network can't resolve that domain, use npx diffthing --offline for plain HTTP on loopback. See how it works.

Agent support

diffthing uses coding-agent CLIs already installed and authenticated on your machine. No provider key is stored by diffthing. Supported: claude, codex, gemini, kimi, qwen, opencode. Force one with --llm; use --llm none for deterministic file-order fallback.

Links

• Source: https://github.com/rahXephonz/diffthing
• License: MIT