Plannotator

Product Introduction

1. Definition: Plannotator is an open-source, local-first interactive plan and code review interface designed specifically for AI coding agents. It serves as a visual feedback layer that integrates with command-line interface (CLI) agents, allowing developers to inspect, annotate, and refine agent-generated plans and code diffs before execution. It functions as a specialized middleware for Agentic Workflows, bridging the gap between raw LLM output and structured human oversight.

2. Core Value Proposition: Plannotator exists to solve the "black box" problem of AI agents by providing a visual UI for human-in-the-loop (HITL) verification. By enabling structured feedback through visual annotations, it ensures AI reliability and safety. Key keywords include AI agent feedback loop, visual plan review, private code review tool, open-source LLM orchestration, and integrated developer environment (IDE) enhancement for Claude Code, Copilot CLI, and Gemini CLI.

Main Features

1. Visual Plan Review & Inline Annotations: Plannotator provides a dedicated browser-based UI that automatically opens when an AI agent completes a planning phase. Users can perform CRUD-like operations on the plan—deleting, inserting, replacing, or commenting on specific steps. These visual changes are translated back into structured feedback that the agent can interpret to revise its strategy.

2. Code Review and Git Diff Analysis: Using the /plannotator-review command, users can visualize git diffs or remote GitHub Pull Requests (PRs). The tool packages these code changes into an interactive format where users can provide line-by-line feedback. This feature includes "Ask AI" capabilities, allowing users to query an agent about specific code blocks during the review process.

3. Versatile Content Annotation (/plannotator-annotate): This feature allows users to ingest and mark up various data types—including folders, URLs, documents, markdown files, and recipes. It is highly compatible with personal knowledge management systems, such as creating a Karpathy-style "LLM Wiki" within Obsidian. Users can annotate the last message sent by an agent using /plannotator-last to provide immediate corrective feedback.

4. Privacy-First Zero-Knowledge Sharing: Plannotator prioritizes security for enterprise and individual use. Small plans are stored entirely within the URL hash (no server storage). Large plans are handled via an end-to-end encrypted short-link service. Data is encrypted using AES-256-GCM in the browser before upload; the server only sees ciphertext, and the decryption key remains solely in the user's shared URL.

Problems Solved

1. Pain Point: Lack of granular control over AI agent execution. Traditional CLI-based agents often present "all-or-nothing" execution paths. Plannotator solves the "hallucination risk" by allowing developers to intercept and correct faulty logic at the planning stage rather than debugging broken code after the fact.

2. Target Audience: Software Engineers utilizing AI coding assistants (Claude Code, Copilot, Gemini), DevOps professionals managing automated infrastructure-as-code (IaC) agents, Technical Leads performing PR reviews, and researchers building "Second Brain" systems or LLM Wikis in Obsidian.

3. Use Cases:

• Complex Refactoring: Reviewing a 20-file plan generated by an agent to ensure architectural patterns are followed.
• Pull Request Auditing: Using /plannotator-review <pr-url> to visually inspect and comment on team contributions.
• Knowledge Synthesis: Annotating long-form documentation or books to create structured datasets for agentic retrieval.

Unique Advantages

1. Differentiation: Unlike built-in agent interfaces which are often text-heavy and ephemeral, Plannotator provides a persistent, visual, and shareable environment. It is vendor-agnostic, supporting a wide array of agents (Claude, Copilot, Gemini, Pi, Codex, OpenCode) rather than locking users into a single ecosystem.

2. Key Innovation: The dual-licensing (MIT/Apache-2.0) and local-first architecture ensure that sensitive proprietary code never leaves the local environment unencrypted. Its "Zero-Knowledge" sharing mechanism is a significant advancement over standard pastebins, making it suitable for secure team collaboration on AI-generated plans.

Frequently Asked Questions (FAQ)

1. Is Plannotator private and secure for enterprise code? Yes. Plannotator runs locally on your machine. For plan sharing, it utilizes AES-256-GCM encryption where the decryption key never touches the server. The server stores only encrypted ciphertext, ensuring a zero-knowledge architecture.

2. Which AI agents are compatible with Plannotator? Plannotator currently supports a broad spectrum of agents including Claude Code, GitHub Copilot CLI, Gemini CLI, OpenCode, Pi, and Codex. It integrates via a plugin architecture or simple slash commands.

3. Can I use Plannotator to build an LLM Wiki in Obsidian? Absolutely. By using the /plannotator-annotate command on markdown files or folders, you can generate structured annotations and feedback that are easily integrated into Obsidian, following the "LLM Wiki" methodology popularized by Andrej Karpathy.

4. How do I install Plannotator on different operating systems? For macOS, Linux, and WSL, you can use the curl-based bash installer (curl -fsSL https://plannotator.ai/install.sh | bash). For Windows, a dedicated PowerShell script is available (irm https://plannotator.ai/install.ps1 | iex). After installation, you can add it as a plugin to your preferred agent.