Plannotator

Plannotator is a local review tool that puts a human back in the loop between an AI coding agent and your codebase. It gives you two things a terminal doesn't. First, a proper interface to read and annotate the plan an agent proposes before it runs. Second, a pull-request-style diff viewer for the code that agent writes before any of it gets committed. The line the site leads with sums up the intent, shape what gets built and own what ships.

The problem is that agentic coding moves fast and the approval step is thin. An agent hands you a wall of plan text or a pile of uncommitted changes, and your only real controls are a yes or a no typed into a cramped terminal window. It's tempting to just wave it through, and that's how code you never actually read ends up in your history. Plannotator turns that rushed moment into a real review with the kind of interface you'd expect from a mature code-review tool.

On the plan side you annotate proposals inline, mark deletions, drop comments, and write replacements directly into the agent's plan. Version history tracks every revision and shows diffs between iterations, so when the agent reworks its approach you can see precisely what changed rather than rereading the whole thing. Plans can be stashed in Obsidian or Bear if you like to keep them alongside your other notes. The point is to own the plan before it becomes code, while changing direction is still cheap.

On the code side it's a diff viewer that works without a commit or a pull request. It reads your local working tree across git, jj, and Perforce, shows committed work, uncommitted edits, and untracked files in one place, and lets you leave annotations at the line, token, file, and review level, each with a suggested edit. You get side-by-side or unified diffs, a file tree to navigate large changes, and a per-commit view. When you'd rather review a real pull request or merge request, it pulls those from GitHub and GitLab with filtering, so the same interface covers both agent output and ordinary team review.

What makes this different from the usual pull-request flow is that none of it requires pushing anything first. You can catch a bad decision while the code is still just uncommitted edits in your working tree, before it becomes a commit, a branch, and a review request that three people have to unwind. Reviewing at the token and line level with suggested edits means you're not limited to leaving a comment and hoping someone acts on it, you can write the exact fix you want and hand it straight back to the agent. For anyone reviewing a steady stream of agent output, that tight loop is the difference between genuine oversight and rubber-stamping. It also keeps your entire review history on your own machine, which is what quietly feeds the custom review skills the tool learns over time.

The AI layer is where it leans into the modern workflow instead of fighting it. You can query a diff in plain language and ask what a change does or why it's there. A guided review orders the changeset by importance so you look at the risky parts first. Review agents flag issues with reasoning and a severity level, and over time Plannotator can distill custom review skills from your own annotation history, so it starts catching the specific things you always catch. Your annotations export as markdown straight back into the agent session, which closes the loop without any copy and paste, so the feedback you write becomes the agent's next set of instructions.

It's built for developers and teams who lean on AI agents but refuse to merge code they haven't actually read. It plugs into Claude Code, Codex, OpenCode, Pi, GitHub Copilot, Gemini, and VS Code, and because it also understands plain git, jj, and Perforce, it's useful even on the parts of your work that no agent touched. You invoke it as a review command from inside the agent you're already using, so it slots into your existing habits rather than asking you to adopt a separate app.

Plannotator is free and open source under MIT or Apache 2.0, and it runs entirely on your machine, so plans and code stay local, with encrypted URL-based sharing when you want a teammate to take a look. Installation is a single curl command. A paid Workspaces tier aimed at teams is in the works but currently waitlist-only, which means the tool as it stands today is free to adopt with nothing to buy. If you've ever rubber-stamped an agent's diff simply because reviewing it properly in the terminal was too painful, this is the exact gap it sets out to close.