跪拜 Guibai
← All articles
AI · Open Source · Claude

Plannotator Brings Browser-Based Annotation to AI Coding Agents' Plans and Diffs

By 冬奇Lab ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The chat-based feedback loop for AI coding agents is slow and imprecise; Plannotator turns plan and code review into a structured annotation workflow that attaches to existing agent checkpoints, which changes how teams supervise and correct agent output without leaving the browser.

Summary

Plannotator replaces the slow, imprecise loop of typing feedback into a chat box with a dedicated browser interface that appears at key points in an AI coding agent's workflow. When an agent calls `ExitPlanMode` or finishes a code change, a hook triggers the local service to open a review screen. You can mark up plans section by section, approve or reject with structured comments, and see the diff after the agent revises. Code review works similarly: it pulls git diffs or GitHub/GitLab PR/MRs, and you comment line-by-line before feedback returns to the same agent session.

Beyond plan and code review, the tool includes an Inbox for asynchronous questions and files, built-in AI assistance, Vim keybindings, and integrations with editors like VS Code and Obsidian. Sharing is privacy-conscious: default local storage with no telemetry, small shares go into the URL fragment unencrypted, large Markdown and HTML are encrypted with AES-256-GCM, and the share service can be self-hosted or disabled entirely. Known limitations include an un-disableable GitHub version check, no authentication in remote mode, experimental Codex hooks, and incomplete support for Droid.

The project is a TypeScript Bun monorepo with about 10,000 GitHub stars, dual-licensed under Apache-2.0 or MIT. The author of the coverage notes that while ten agents are listed, integration depth varies from full plan interception to simple commands, so checking the README for your specific agent is advised.

Takeaways
— Plannotator supports ten agents out of the box: Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, DeepSeek Harness, and Pi.
— Plan review triggers automatically via the ExitPlanMode hook (with a 4-day timeout), so a developer can step away and review later.
— Approving a code review sends "LGTM" to the agent; rejecting sends structured comments back to the same session.
— Code review covers local changes and GitHub/GitLab PR/MRs, and integrates with Git, GitButler, Jujutsu, and Perforce.
— Sharing has two tiers: URL-fragment shares are compressed but unencrypted; large Markdown/HTML shares are AES-256-GCM encrypted, expire after 7 days, and the key stays in the URL.
— The GitHub version check on every interface load cannot be disabled, despite the default no-telemetry stance.
— Remote mode listens on 0.0.0.0 with no authentication, so exposing port 19432 to untrusted networks is unsafe.
— The repository is a Bun monorepo with apps/ for per-agent integrations and packages/ for shared UI, editor, server, and core logic.
Conclusions

The design attaches to checkpoints the agent already has, like ExitPlanMode, rather than inventing a new workflow; that's what makes it practical to adopt without restructuring how you use AI coding agents.

Supporting ten agents is less impressive than it sounds because depth varies widely, from full plan interception (Claude Code) to just commands (Droid); the README's per-agent rows are the real integration truth.

The privacy model is deliberately transparent about what leaves the machine, but the un-disableable GitHub version check and the absence of auth in remote mode are quiet trade-offs that a privacy-sensitive team must accept or self-host around.

The author's own caveat that the "no telemetry" claim is unverified because a telemetry directory exists in the repo is a reminder that README statements on privacy deserve source-level scrutiny.

Concepts & terms
Hook
In AI coding agent platforms, a hook is a callback the agent fires when it reaches a specific state, such as requesting permission to run a plan or completing a code change. A local service can listen to these hooks and trigger external actions, like opening a browser UI.
ExitPlanMode
A command or signal in agents like Claude Code that marks the end of the plan-generation phase. When an agent calls it, the platform asks for user permission to proceed, which Plannotator intercepts to show its review interface.
SLSA provenance
Supply-chain Levels for Software Artifacts, a security framework that provides verifiable metadata about how a software artifact was built and who published it. Plannotator ships SLSA provenance for its releases starting with v0.17.2, so users can verify the build chain.
PrivateBin model
A pastebin-style sharing approach where content is encrypted in the browser before upload, the server stores only ciphertext, and the decryption key is embedded in the URL fragment, which never reaches the server. Plannotator uses this for its large-share feature, with a 7-day expiration.
From the discussion
Featured comments
shushenghong

Structured annotations are far more convenient than describing "that line" in the chat box. Another point that's easy to miss: annotations only apply to this session. When the same type of feedback gets annotated a second time, we turn it into a rule and put it in a file in the repo for the agent to read, so every agent in the team will read it first, without each person having to annotate it again.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗