Review your AI agent's plan
before it writes a line of code
When an agent proposes a plan, AgentsRoom opens it in the agent's pane as a document you can read. Select a passage to comment on it, replace it, delete it or add a step after it, then approve the plan or send it back with your notes, in one message.
Claude Code plan mode, Codex Plan mode, the OpenCode Plan agent and Antigravity's implementation plan, on macOS, Windows and Linux, and in the AgentsRoom app on your phone.
Reading the auth module and the sessions table
Plan ready, waiting for your review
Add a password reset
- 1.Create a new table for reset tokensCommentReplaceInsert afterDelete
- 2.Email a link that works only once
- 3.Rewrite the whole auth moduleCommentReplaceInsert afterDelete
- 4.Expire the link after 30 minutes
Your annotations
ReplaceCreate a new table for reset tokens → Reuse the sessions table
DeleteRewrite the whole auth module
General note
Keep the current login flow as it is.
The plan arrives as a document, you mark what has to change, and one click sends it all back to the agent.
Plan mode exists so you can stop a wrong approach before it turns into an hour of agent work. In a terminal, that review is the hard part: the plan scrolls by in monospace, the prompt under it asks you to approve or keep planning, and any correction has to be typed from memory, quoting the step you mean.
AgentsRoom turns that moment into a real review. The plan opens as a rendered document in the agent's own pane, right above its input box, where the agent's questions already show up. You read it like a spec, mark the passages you disagree with, and decide: approve, approve with notes, or request changes.
Your annotations leave as one structured message: each passage quoted, what to do with it, then your general note. The agent revises and proposes a new version, and the Changes tab shows exactly what moved since the previous one.
What the plan card does
What you need to review a plan properly, in the place where the agent already waits for you.
A plan you can actually read
Headings, lists, tables and code blocks rendered as a document instead of scrolling by in monospace. Enlarge the card for a long plan, fold it to glance at the terminal.
Comment on a passage
Select any text and a small menu opens under it. Comment says what worries you about that exact passage, and the passage stays highlighted.
Replace it
Write what the passage should say instead. The agent receives the original and your version, so nothing is left to interpretation.
Add what is missing
Insert after adds a step, a check or a constraint right where it belongs, after the passage you selected.
Cut the scope
Delete marks a step the agent should drop. Handy when the plan quietly grew a refactor nobody asked for.
A general note
For what no single passage covers: a priority, a file not to touch, a rule for the whole change. It is optional.
One structured message
Annotations and note leave together: each passage quoted, one line per annotation, written in your language. You can remove an annotation before sending.
Every version, with a diff
The card numbers each version of the plan: v1, v2, v3. The Changes tab shows what changed since the previous one, line by line.
How a plan review works
Four steps, and no code is written before you say so.
The agent proposes a plan
You start the agent in plan mode, or it reaches its plan step on its own: Claude Code plan mode, Codex Plan mode, the OpenCode Plan agent, an Antigravity implementation plan.
The plan opens above its input box
A card titled Proposed plan appears in the agent's pane, with the version number of the plan. Nothing pops over your screen: the terminal stays right there.
You annotate
Select a passage, pick Comment, Replace, Insert after or Delete, type, press Enter. Do it as many times as you need, then add a general note if you want.
You decide
Approve, Approve with notes, or Request changes. With changes, the agent stays in plan mode, revises the plan, and its next version opens in the same card.
Four ways to answer
Each one is a single click, and each one says exactly what the agent does next.
Approve
The agent leaves plan mode and implements the plan as written.
Approve with notes
Offered as soon as you have written something. The plan is approved and your notes go with it, for the agent to apply while it implements.
Request changes
Your annotations and your note go back to the agent, which revises the plan and proposes it again before writing any code.
Answer in the terminal
The card closes and the CLI's own prompt takes over, for the times you would rather type your answer there.
One card for four agent CLIs
Each CLI has its own plan step. AgentsRoom meets each one where it is, and the card stays the same.
Claude Code
Plan mode. The agent waits for your answer in the card: approving leaves plan mode, requesting changes keeps the agent planning with your feedback. It waits as long as you need, up to 24 hours.
Codex
Plan mode. When a turn ends with a proposed plan, AgentsRoom opens the same card. Your answer is typed in the terminal for you: the menu choice when you approve, your annotations as a message when you request changes.
OpenCode
The Plan agent. AgentsRoom spots the plan at the end of the turn, opens the card, and types your answer in the terminal for you.
Antigravity
The implementation plan the agent submits for review. AgentsRoom picks it up at the end of the turn and opens it in the same card.
Every other agent CLI
No plan step to catch? Annotate the last message gives you the same card on the agent's latest reply, with all 17 agent CLIs AgentsRoom supports (Claude, Codex, GitHub Copilot CLI, Cursor and 13 more).
Nothing to install. For Claude Code, AgentsRoom sets up the hook that lets the agent wait for your answer. In Codex, OpenCode and Antigravity it installs no hook at all: the plan is spotted when the turn ends and your answer is typed for you, the way you would type it.
What a plan review looks like
Real annotations, the kind that save an hour of agent work each.
Step 3: rewrite the auth module
The plan grew a refactor nobody asked for. One deletion, and the ticket stays the size it was.
Create a new table for reset tokens → Reuse the sessions table
You know the schema better than the agent's first look at it.
Run the migration → Then stop and wait for my go before production
The step the agent could not guess: how you ship.
Change the format of the public API response
Ask why before it happens: two client apps depend on that format.
Do not touch the billing module in this change.
A rule for the whole plan, not for one passage.
Call the setting Password reset, not Account recovery.
The plan is right, and one detail is not worth a new version.
Not only plans: any reply, any spec
The same card reviews what the agent just told you, or a markdown file you want an agent to act on.
The agent's last message
Right-click in the agent's terminal and choose Annotate the last message. Its reply opens as a document: comment, replace, delete or add, then Send to the agent. Works with every agent CLI.
A spec or a plan file
Open a markdown file in the file preview, click Annotate, mark it up, then pick the agent of the project that receives your notes.
On a message or a file, your notes arrive as one message, like a message typed in the agent's input box.
Every version, on every screen
A plan review does not stay stuck on the machine where it started.
History on your account
Every version of the plan in the conversation is kept: in the Changes tab of the card, and on every computer signed in to your AgentsRoom account.
Review from your phone
The card shows up in the agent's console in the AgentsRoom app for iOS and Android. Tap a paragraph to comment on it, replace it, delete it or add text after it, then approve or request changes. Your paired computer applies the answer.
It waits for you
With Claude Code the card waits up to 24 hours. If AgentsRoom is closed or restarted in the meantime, nothing is decided in your place: the CLI shows its own approval prompt.
On by default, off in one switch
Plan review is on for every agent. To go back to each CLI's own approval prompt, turn off Review plans in AgentsRoom in Settings > Agent behavior. Pressing Esc in the agent's terminal also closes the card.
Frequently asked questions
How do I review a Claude Code plan before it writes code?
Run Claude Code in AgentsRoom and put it in plan mode, or ask it for a plan. When it proposes one, the plan opens in a card above its input box. Select passages to comment on them, replace them or delete them, then click Approve or Request changes. Plan review is on by default.
Does it work with Codex Plan mode?
Yes. When a Codex turn ends with a proposed plan, AgentsRoom opens the same card. Your answer is typed in the terminal for you: the menu choice when you approve, your annotations as a message when you request changes, and Codex stays in Plan mode to revise.
Which agent CLIs does plan review support?
Claude Code, Codex, OpenCode and Antigravity, the CLIs whose plan step AgentsRoom catches. Annotate the last message works with all 17 agent CLIs AgentsRoom supports (Claude, Codex, GitHub Copilot CLI, Cursor and 13 more), so any agent's reply can be reviewed the same way.
What does the agent receive when I request changes?
One message: a line saying you reviewed the plan and want changes, each annotated passage quoted with what to do with it (your comment, its replacement, the text to insert or a deletion), your general note, and a request to revise the plan and propose it again before implementing. It is written in your language.
Can I still answer in the terminal?
Yes. Answer in the terminal closes the card and hands the decision back to the CLI's own prompt. To skip the card entirely, turn off Review plans in AgentsRoom in Settings > Agent behavior.
Where do I find the previous versions of a plan?
In the Changes tab of the card, which shows line by line what changed since the previous version, and on every computer signed in to your AgentsRoom account.
Can I review and approve a plan from my phone?
Yes. The card appears in the agent's console in the AgentsRoom app for iOS and Android: tap a paragraph to comment on it, replace it, delete it or add text after it, then approve or request changes. Your paired computer applies the answer. Very long plans are cut at 60,000 characters on the phone, and the desktop keeps all of it.
What happens if I do not answer for hours?
The card waits. With Claude Code it waits up to 24 hours. If AgentsRoom is closed or restarted in the meantime, nothing is decided in your place: the CLI shows its own approval prompt instead.
Can I annotate something that is not a plan?
Yes. The agent's last message, from the right-click menu of its terminal (Annotate the last message), and any markdown file opened in the file preview (Annotate), sent to the agent of your choice. Your notes leave as one message.
Do I have to install a plugin or a hook?
No. For Claude Code, AgentsRoom sets up the hook that lets the agent wait for your answer. Codex, OpenCode and Antigravity get no hook at all: the plan is spotted at the end of the turn and your answer is typed in the terminal for you.
Is plan review free?
Yes. It is part of every AgentsRoom plan, the free one included, on macOS, Windows and Linux and in the mobile app.
Further reading
Should You Still Review Your AI Agent's Code?
Your agents write better code than half the pull requests you used to merge. So do you still read every line? The honest case for both sides, the 10 signals that tell you an agent messed up, and how much review each change actually deserves.
How to Run 3 to 8 Coding Agents in Parallel Without Losing Track
Launching multiple Claude Code or Codex agents at once is easy. Keeping control is where it falls apart. Here's the method that actually works.
30 Hook Events Fire in a Claude Code Session. Only 3 Can Talk Back.
The complete list of Claude Code hook events, when each one fires, which 15 can block, and the stdout rule that silently swallows most hook output. A field reference built from running hooks in production across thousands of agent sessions.
Related features
Structured questions
Your AI coding agent asks in a card with clickable options, right above its input box, and waits for your answer. Codex, Claude Code, Cursor, Copilot CLI.
Review Mode
Review every file your AI agents modified with a side-by-side diff viewer, syntax highlighting, word-level diffs, keyboard navigation, and inline comments sent back to the agent. Commit when ready.
Per-Agent Review
Review code changes from each AI agent separately. AgentsRoom tracks which Claude Code agent modified which files. Filter diffs by agent, review individually, commit per agent for clean git history.
Remote Control
Remote control your Claude, Codex and Antigravity coding agents from iPhone or Android. Pilot terminals, drive projects, operate the backlog and tunnel dev servers. Exclusive encrypted relay, built into AgentsRoom.
Ticket scoping
Ticket Scoping frames a vague feature request into a clear, validated need before any code is written. One click on a feedback ticket spawns a Product Manager AI agent that reads your codebase and builds an HTML mockup faithful to your real product, with the requested change in context. Review the mockup, share it with the reporter, and keep clarifying in a two-way conversation on the same ticket. No task started, no notification sent. Works with Claude, Codex and Antigravity.
Plannotator alternative
Plannotator reviews AI agent plans in a browser tab. AgentsRoom opens them in the agent's own pane, on your computers and your phone. An honest comparison.
Review the next plan before it becomes code
Download AgentsRoom, start an agent in plan mode, and the plan opens in its pane. Free for 2 projects.
Companion app: monitor your agents on the go
Bring your own: Claude, Codex, Antigravity CLI, or other AI provider.
Push bugs and requests straight to your public backlog.