all projects

live nittle.dev

nittle

Review your agent's diff like a PR, before there is a PR, and hand the comments straight back.

Visit nittle.dev Source Install
nittle in dark mode reviewing a working tree: a file tree on the left, a TypeScript diff with flagged changes, a collapsed generated file, and a block marked moved from router.ts

What it is

A local git diff viewer for reviewing what coding agents write. Comment on lines, ranges or whole files like a PR, attach a suggested change, then return the feedback to the agent over a local MCP server, by paste, or on stdout with nittle --wait. The agent replies in threads, marks comments addressed and asks for a re-review, and comments from earlier rounds carry over. A deterministic scorecard and review order (no AI involved) tell you where to look first, moved code is linked even across files, and go to definition works inside the diff for TypeScript, C#, Python, Go, Rust, Vue and Svelte. One MIT licensed binary for macOS, Linux and Windows that only needs git, and nittle setup wires it into Claude Code, Codex, Cursor and six other agents. It started as loupe in June 2026 and was called diffle for a while.

Stack

  • TypeScript
  • Bun
  • Svelte
  • Vite
  • Tailwind CSS
  • MCP
  • LSP

Screenshots

  • nittle in light mode with a comment on a rate-limit return in src/server.ts and a suggested change that adds a Retry-After header, rendered as a diff
  • nittle's review scorecard: size low, tests high with five untested files listed, public API high, dependencies and hotspots medium, flags high
  • nittle showing hover docs for createLimiter on an added line, pulled from the TypeScript language server

Writing about nittle

  1. AI & agents

    445 Commits in 32 Days: nittle v0.24 to v0.34

    Between September 29 and October 30, nittle shipped 14 releases, picked up code navigation and a vim keymap, got renamed for the second time in a month, and roughly quadrupled its tests. Here's what landed and why it matters.

  2. AI & agents

    Why I Built nittle

    I got tired of reviewing my agent's work in a PR and copying every comment back by hand, so I built a local diff viewer that sends the review straight to the agent. Five months and two renames later, it's nittle.

  3. AI & agents

    Revisiting Autonomous Development with Opus 4.8

    I handed Claude Opus 4.8 a build spec and got back a working, tested, published tool in one session. Here's the framework that made a long autonomous run reliable instead of reckless.