backnotprop/plannotator

If you already run a terminal coding agent and mostly read its plans and diffs in a scroller, Plannotator changes that step: proposals and changes open in a browser page you mark up, and your comments go back to the agent as input rather than being retyped.

The problem

A typical loop with Claude Code, Codex or Gemini CLI looks like this: the agent proposes a plan, you scroll through it in the terminal, then type corrections in prose. For diffs, you either skim them in the terminal or switch to a separate code review tool. Nothing links your line-level thoughts to the agent that will implement them. Plannotator inserts that link: the plan, spec, HTML artifact or diff opens in a local browser surface, you comment inline or on specific lines, and the batch of feedback is handed back for the agent to act on.

What's different

The core idea is the round trip. Annotation is not a separate document you keep for yourself; it becomes agent input. That is why the integration is hook- and command-based rather than a standalone app you open on the side, and why the same surface covers plan review, HTML artifacts and code review work, including local changes and GitHub/GitLab pull requests, with GitButler, Jujutsu and Perforce named among the supported version-control setups. A shared review and inbox concept extends this to team or multi-agent threads.

The trade-off is that it wraps your agent's workflow instead of sitting beside it. The installer auto-detects installed agents and configures hooks, skills and slash commands, so the comfortable path assumes a host agent that supports that plumbing. There is also a --minimal path that installs only the binary to ~/.local/bin, but then you finish per-agent setup yourself. Navigation and review quality depend on the browser tab the tool opens, not an IDE.

The privacy model is worth reading rather than assuming. Plans, diffs and annotations stay local by default, but each review surface checks GitHub for the latest release on load with no opt-out, and local Git review can query the configured origin on certain actions. That is not a content leak, but it is traffic some teams will want to know about. Remote checks can be disabled with a flag, environment variable or config setting. Network features (URL annotation via Jina Reader, PR retrieval through your authenticated CLI) are the points where content leaves the machine.

Who should try it

Good fit: a developer who already works through plans with a terminal agent, wants comment-on-the-plan and comment-on-the-diff as one habit, and is comfortable letting the installer write hooks and commands into their agent config. It also suits reviewers who want a browser-based diff with line comments that feed back to the coding agent.

Poor fit: someone who wants a general-purpose, agent-agnostic code review tool that runs outside the agent loop, or who needs the install to touch only a single binary with no host configuration. The minimal install avoids that but shifts the setup burden to you. Readers on sandboxed setups should expect extra configuration: one commenter reported getting it working inside a sandbox only after setting PLANNOTATOR_REMOTE=1 and allowing the app's directory and port, which illustrates that environment-specific plumbing may be needed.

Requirements and limitations

Setup is the main obvious cost: one installer covers macOS, Linux, WSL and Windows, then a per-agent finishing step is required, and detailed instructions live in the installation docs. The installer may contact GitHub, npm and other services depending on options, so offline or tightly egress-controlled machines need the minimal path plus manual wiring.

Ongoing upkeep matters too. Releases ship with SHA256 sidecars and SLSA provenance from v0.17.2; the SBOM is explicitly a monorepo-wide input inventory, not a per-binary runtime list, so it is not a drop-in audit of exactly what a given binary bundles. An ordinary uninstall keeps plans and history; full removal requires typing a confirmation word. Host integrations are always part of uninstall, which can block cleanup if an agent's plugin manager is broken.

HotRepo's take

HotRepo's records are thin: on the daily board it appeared on 2026-10-24 at 9,614 stars (+24) and 2026-10-25 at 9,617 (+20), with no comparable history before that, so persistence cannot be judged from these two snapshots. The decision point is not popularity but fit: check the privacy and network behavior and whether your agent is on the installer's detection list before adopting, since the hook-and-command wiring is what makes or breaks the workflow.