skulldc

Plannotator gives coding agents a local browser surface for plan and code review

The Apache-2.0 Plannotator project hooks into agents such as Claude Code, Codex and Gemini CLI so plans, documents and diffs can be annotated and sent back as feedback.

Plannotator, the backnotprop/plannotator repository on GitHub, is a local, browser-based review surface for AI coding agents. Its README lists Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, DeepSeek Harness and Pi as supported agents. The project is written in TypeScript, licensed Apache-2.0 and was created on 2025-12-28. At the time of the trending snapshot it had 11747 stars, 860 of them gained in the last day.

How it plugs into an agent

The README says Plannotator connects to an agent through plugins, hooks and commands. When the agent proposes a decision, a plan or HTML, or finishes working, the app opens in the browser. The user marks the content up, comments, and sends the feedback directly to the agent to act on. Plan mode needs no command: it is wired in through each harness hook, so the markdown review surface opens any time the agent creates a plan.

Code review works on local changes or remote PRs. Comments on diffs and suggested code go back to the agent, and the README lists Git, GitButler, Jujutsu, Perforce, GitHub and GitLab as supported. AI is built in as well: a reviewer can ask AI about what is being reviewed or launch AI reviews that post comments to the diff.

The newer Plannotator Inbox is one local window where agents leave messages, questions, files and guided reviews while they keep working. According to the README, nothing holds an agent session open while the user answers, and the reply goes back to the session that asked. Claude Code, Pi and OpenCode connect through plugins; other agents can reach the Inbox over MCP with plannotator inbox mcp.

Installation

One installer places the plannotator binary, detects installed agents and configures hooks, skills and slash commands for each:

curl -fsSL https://plannotator.ai/install.sh | bash

Passing –minimal installs only the binary to ~/.local/bin and skips every skill, hook, slash command and per-agent config. Some agents need one more step afterwards; for Claude Code the README lists adding the plugin marketplace and installing the plugin, then restarting. On Claude Code 2.1.287 or newer, plan review, code review and annotate no longer make Claude wait: the turn ends and the decision arrives later as a message.

Network behavior and sharing

The README documents its network boundaries in some detail. Plannotator does not collect usage telemetry or analytics, and plans, diffs, annotations, drafts, history and configuration stay local by default. Each app surface does check GitHub for the latest Plannotator release when it loads; that sends no plan or review content, but there is currently no opt-out setting for it. Local Git review can query the configured origin with git ls-remote to detect the default branch; that can be turned off with –no-git-remote-check or PLANNOTATOR_GIT_REMOTE_CHECK=0.

Other features send content out only when they need to. URL annotation fetches pages through Jina Reader by default, and Ask AI sends the selected question and relevant context to the configured provider.

Link sharing has two modes. Small markdown shares are compressed into the URL fragment, which is not encrypted, so anyone with the complete link can read the content. Large markdown and raw HTML shares are encrypted with AES-256-GCM before upload, the server stores only ciphertext, the key stays in the URL fragment, and hosted pastes expire after 7 days. Sharing can be disabled with PLANNOTATOR_SHARE=disabled. The README adds that open source link sharing is moving to deprecated support, with Workspaces, a separate hosted product, as the primary direction for team sharing. No removal date has been announced.


Source: backnotprop/plannotator on GitHub