Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitHub Trending AI→ original

Plannotator Lets You Comment on Plans from 10 AI Agents in the Browser: 100 GitHub Stars in a Day

When an AI agent returns a thirty-step plan or edits twenty files, replying to it with a paragraph in the terminal is awkward. Plannotator opens the plan or diff in the browser itself: you comment on the lines you need, hit Send, and your remarks go back to the agent. It works with 10 agents, including Claude Code and Codex, and gained 100 stars on GitHub in a day. That makes reviewing agent work feel like a normal pull request review.

AI-processed from GitHub Trending AI; edited by Hamidun News
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitHub Trending AI ↗
Editor-in-Chief
All data stays localPlannotator's server does not store plans or history
One-command install via install.shit finds your agents and installs their hooks
Key metric #3/plannotator-review opens a diff or PR, /plannotator-last opens the agent's last response

Imagine you asked an AI agent to finish a project, and it returned a thirty-step plan or edits across twenty files. Reading that in the terminal and replying with one long paragraph is awkward, and half the comments get lost. Plannotator solves this problem. It is an open-source project from backnotprop. It opens the agent's plan or diff in the browser, where you comment on specific lines and send your feedback back. The tool works with 10 coding agents, from Claude Code to Codex, and gained about 100 stars on GitHub in a single day, landing on the trending list.

What it looks like in practice

The authors describe Plannotator as the missing review layer between a human and multi-agent work. It is a local browser surface that connects to the agent through plugins, hooks, and commands. When the agent proposes a solution, drafts a plan, or finishes its work, the interface opens in the browser automatically. You mark up the text, leave comments, and approve or reject. If you approve, the agent continues. If you reject, it receives structured feedback, revises the plan, and the interface shows what changed between versions.

Supported agents include Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, DeepSeek Harness, and Pi. A separate Plannotator Inbox also works with personal assistants: Grok Bot, Muse, Hermes, and OpenClaw. It is a single local window where agents leave messages, questions, files, and analyses while they keep working. You can answer a question with one click. The agent's session does not hang while waiting, and your reply arrives as its next turn. Other agents connect via MCP.

What you can comment on

First, documents: markdown, specifications, HTML, entire folders, and URLs. The /plannotator-annotate command accepts a path or address, and /plannotator-last opens the agent's latest message.

Second, code. /plannotator-review shows uncommitted changes, or a GitHub pull request or GitLab merge request by link. Git, GitButler, and Jujutsu are supported. You can comment on lines, suggest edits, and send everything to the agent. The agent can also split a large diff into "chapters": first the essence of the change, then its consequences, then supporting code. Each chapter has a short description next to the actual diff.

Third, live sites. The command plannotator annotate http://localhost:5173 opens your local app through a proxy, so you comment on the real page while hot reload keeps working. For the terminal there is Herdr Annotate: it marks up console text and agent responses, and annotations are saved to Plannotator's shared data directory.

You can send a plan to a colleague as a link: they leave notes, and you import them into your agent. Small materials live directly in the link, while large ones are encrypted before upload and deleted after 7 days. Anyone with the full link can read the content. Sharing can be turned off with the environment variable PLANNOTATOR_SHARE=disabled. Approved plans can be automatically saved to a notes vault with tags and backlinks, or to Bear or Notion.

Installation and limitations

A single installer (curl with install.sh) downloads the binary from GitHub Releases, detects installed agents, and sets up hooks, skills, and slash commands for each one. If you only want the binary, you can tell the installer not to touch agent configs. Claude Code and Copilot CLI have plugins from the marketplace, OpenCode and Pi have separate extensions, and terminal mode is installed with brew install plannotator/tap/plannotator-tui.

Data is stored only locally, Plannotator's server has none of it, and it cannot be recovered remotely. Uninstalling by default removes the installed components but keeps your plans, history, and settings. A full wipe requires the --purge flag, and with non-standard paths or broken configs the command stops and explains what to do manually.

There are caveats:

  • In Codex, hooks are still experimental, and on native Windows the installer prints manual setup instructions.
  • For Droid, only commands are available for now, without plan interception.
  • Gemini CLI must be version 0.36.0 or newer.
  • In SSH sessions Plannotator switches to a fixed port, while in devcontainers and Docker the port is set separately.

Why it matters for practitioners

Reviewing an agent's work takes the same shape as reviewing a colleague's: comments on specific lines instead of long chat replies. If you run several agents at once, the single Inbox saves you from hopping between terminals. In Claude Code version 2.1.287 and newer, plan review, code review, and annotation do not make Claude wait: it ends its turn, and your decision arrives later as a message.

To test the idea quickly, install the tool and call /plannotator-last in your usual agent. From then on, the plan opens in the browser automatically as soon as the agent proposes it.

Story

Key Questions Answered

What it looks like in practice?

The authors describe Plannotator as the missing review layer between a human and multi-agent work. It is a local browser surface that connects to the agent through plugins, hooks, and commands. When the agent proposes a solution, drafts a plan, or finishes its work, the interface opens in the browser automatically. You mark up the text, leave comments, and approve or reject. If you approve, the agent continues.

What you can comment on?

First, documents: markdown, specifications, HTML, entire folders, and URLs. The /plannotator-annotate command accepts a path or address, and /plannotator-last opens the agent's latest message. Second, code. /plannotator-review shows uncommitted changes, or a GitHub pull request or GitLab merge request by link. Git, GitButler, and Jujutsu are supported. You can comment on lines, suggest edits, and send everything to the agent.

What are the key details on Installation and limitations?

A single installer (curl with install.sh) downloads the binary from GitHub Releases, detects installed agents, and sets up hooks, skills, and slash commands for each one. If you only want the binary, you can tell the installer not to touch agent configs. Claude Code and Copilot CLI have plugins from the marketplace, OpenCode and Pi have separate extensions, and terminal mode is installed with brew install plannotator/tap/plannotator-tui.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…