Diffsmith

Product Introduction

1. Definition: Diffsmith is a native macOS desktop application, specifically a local-first code review studio designed for analyzing uncommitted git diffs. It operates as a technical intermediary between a developer and their AI coding assistant.
2. Core Value Proposition: It exists to provide a structured, offline environment for developers to conduct precise, line-anchored code reviews on changes generated by AI agents like Claude Code, Cursor, or GitHub Copilot before those changes are committed, enabling faster, more accurate iteration cycles.

Main Features

1. Real-time Git Diff Analysis: Diffsmith continuously monitors the local git working tree, displaying syntax-highlighted diffs (side-by-side or inline) against HEAD or HEAD plus staged changes. It automatically updates the view as files change on disk, allowing review to begin before an AI agent finishes its task.
2. Inline Commenting with Context Anchoring: Users can click any line in the diff to attach a comment directly to that specific code location. This eliminates manual tracking of file names and line numbers. All comments are listed in a navigable sidebar, and comments are intelligently moved to a "Past Comments" section if the underlying code changes.
3. AI-Agent Integration via Prompt Export and MCP: Reviewed comments can be compiled and copied as a formatted, AI-ready prompt for pasting into any chat interface. Alternatively, Diffsmith's built-in, local-only Model Context Protocol (MCP) server allows compatible agents (Claude Code, Cursor, etc.) to read comments directly and post replies anchored to the relevant code within the Diffsmith interface, creating a closed feedback loop.
4. Offline-First, Privacy-Centric Architecture: The application is a sandboxed native macOS app with no telemetry, no cloud account, and makes zero outbound network calls. All processing occurs locally. The optional MCP server is strictly loopback (localhost), ensuring code and comments never leave the user's machine unless explicitly shared via the AI agent's own channels.

Problems Solved

1. Pain Point: The disjointed workflow of reviewing AI-generated code, which typically involves switching between an IDE/terminal (to see the diff), a notepad (to jot notes), and the AI agent's chat interface (to relay feedback), often resulting in miscommunication due to imprecise line references.
2. Target Audience: Software developers and engineers who regularly use AI-powered coding assistants (Claude Code, Cursor, GitHub Copilot, Amazon CodeWhisperer) within their IDE and require a systematic way to audit and refine the agent's output. Particularly valuable for senior developers reviewing junior or AI-generated code and for solo developers seeking higher-quality AI assistance.
3. Use Cases: 1) Conducting a security and logic review on a complex refactor suggested by an AI. 2) Providing style guide and optimization feedback on a new feature generated by Claude Code. 3) Using the MCP integration to have the AI agent explain its reasoning for a specific block of code directly in the review context. 4) Batch-reviewing all changes made during a lengthy AI-assisted debugging session before deciding what to commit.

Unique Advantages

1. Differentiation: Unlike cloud-based pull request review tools (GitHub, GitLab, LinearB), Diffsmith operates on uncommitted, local changes. Unlike generic diff tools (Delta, Kaleidoscope), it is purpose-built for the AI-agent workflow with integrated commenting and agent feedback loops. It is not a subscription service but a one-time purchase.
2. Key Innovation: The integration of a local MCP server transforms the product from a passive review tool into an active collaboration platform. This allows the AI agent to become a participant in the review conversation within the context of the diff, which is a significant technical and UX advancement over manual copy-paste workflows.

Frequently Asked Questions (FAQ)

1. Is Diffsmith compatible with GitHub Copilot and JetBrains IDEs? Yes, Diffsmith is agent-agnostic. It reviews the git diff in your local repository, so it works with any AI coding tool that writes to your filesystem, including GitHub Copilot (in VS Code, JetBrains IDEs, etc.), Amazon CodeWhisperer, and any LLM used via CLI. Feedback is delivered via copied prompt or MCP.
2. How does the MCP integration work and is it secure? Diffsmith runs a local MCP server only on your machine (localhost). You enable it in the app and configure your AI agent (following provided snippets) to connect to this local server. All communication is loopback; your code and comments are never transmitted to Diffsmith's servers, as none exist.
3. Can I use Diffsmith to review my own code, not just AI-generated code? Absolutely. While optimized for the AI-agent workflow, Diffsmith functions as a powerful, offline git diff viewer and comment tool for any uncommitted changes, making it useful for reviewing your own work or that of a human collaborator via patch files.
4. What happens to my comments if I close the app or the repo? Comments are stored locally on your Mac. When you re-open the same repository, Diffsmith will reload your previous comments, anchoring them to the current code. If the code has changed significantly, comments may be moved to the "Past Comments" section for reference.
5. Does the one-time purchase include future updates? Yes, purchasing Diffsmith on the Mac App Store grants lifetime access to the application, including all future updates and feature additions for macOS. There is no subscription or recurring fee.