Plannotator: Review AI Agent Plans Before They Run (Open Source) Plannotator, an open-source local workspace from backnotprop/plannotator that intercepts AI coding agents at the plan stage for review before execution, reached 11,523 GitHub stars on November 15 after gaining 860 stars in 24 hours. The MIT/Apache 2.0 tool installs a hook that intercepts Claude Code's ExitPlanMode, opening plans in a browser for approval or annotation, and works with nine agents including Codex, Copilot CLI, Gemini CLI, OpenCode, Pi, Amp, Droid, and Kiro CLI. The v0.27.19 release added review of changed images in diffs and real GitHub file-thread comments, while v0.25.0 drew 18 pull requests, 13 from outside the core team and 8 from first-time contributors. AI coding agents write code faster than any human reviewer. That speed is the problem. Plannotator is a local, open-source workspace that intercepts your agent before it executes, presents you with a visual plan to review and annotate, and returns your feedback directly into the agent’s context. It just hit GitHub Trending with 860 stars in 24 hours — and if you run Claude Code, Codex, or Copilot CLI, it fills a gap you probably didn’t know you had. The Problem With Agent-Speed Development When AI coding agents move faster than your ability to review, the review step gets skipped — or becomes a rubber stamp. Research on human oversight of agentic systems https://arxiv.org/pdf/2606.05391 flags the failure mode clearly: the faster agents write, the longer review queues grow, and human oversight degrades into a formality performed under time pressure. Plannotator takes a different approach. Rather than reviewing after a PR is open, it intercepts the agent at the plan stage — before a single file changes. What Plannotator Does Plannotator backnotprop/plannotator https://github.com/backnotprop/plannotator on GitHub is a free, browser-based workspace for reviewing AI agent plans, annotating code diffs, and pushing structured feedback back into agent workflows. It runs entirely locally — no cloud, no data leaving your machine. License is MIT or Apache 2.0, your choice. - Plan Review — Visual UI that opens when your agent finishes planning. Approve the plan or annotate it and send feedback for revision. - Code Review — Diff viewer for local changes, GitHub PRs, and GitLab MRs. Comment on specific lines, add suggestions, and return the result to your active agent session. - Inbox — Asynchronous message center where agents leave questions, files, and guided reviews for you to handle on your own schedule. How the Plan Review Hook Works In Claude Code, Plannotator installs a hook that intercepts ExitPlanMode . When the agent finishes planning, a local server starts, reads the plan, and opens it in your browser. You can approve agent proceeds or annotate and send feedback — the agent revises its plan and loops back. On Claude Code 2.1.287 and later, review is non-blocking: Claude ends its turn and your decision arrives as a follow-up message, so there is no timeout pressure. Setup for Claude Code: Install Plannotator curl -fsSL https://plannotator.ai/install.sh | bash Add Claude Code plugin /plugin marketplace add backnotprop/plannotator /plugin install plannotator@plannotator Alternatively, use the hook script: bash scripts/setup-hook.sh — it merges the ExitPlanMode hook into ~/.claude/settings.json without overwriting your existing configuration. Code Review Before the Commit Most AI code review tools — CodeRabbit, Qodo, Greptile — review after a PR is already open. Plannotator reviews local diffs before any commit, and your line-level comments go straight back to the active agent session. It supports Git, Jujutsu, GitButler, and Perforce locally, plus GitHub PRs and GitLab MRs by URL. The v0.27.19 release added support for reviewing changed images in diffs and posting real GitHub file-thread comments. Beyond Claude Code, it works with Codex, Copilot CLI, Gemini CLI, OpenCode, Pi, Amp, Droid, and Kiro CLI — nine agents in total, through host-specific hooks, plugins, or skills. The official documentation https://docs.plannotator.ai/open-source/workflows/plan-review covers per-agent setup in detail. Community Momentum Plannotator reached 11,523 GitHub stars https://www.tokenpost.com/news/technology/57781 on November 15, gaining 860 in a single day. The v0.25.0 release was the project’s largest community release: 18 pull requests, 13 from outside the core team, 8 first-time contributors. This is not a solo developer’s side project. It is becoming a standard fixture in AI-augmented development workflows. The Missing Layer The prevailing assumption is that the right place to review AI-generated code is the pull request. Plannotator challenges that. By the time a PR is open, the agent has already made dozens of decisions without human input. Plan review catches those decisions before they happen. For teams that want to stay in control of their codebase without slowing agents down, that is the right intervention point. Source code and install instructions at plannotator.ai https://plannotator.ai/ .