An async surface for your long‑running agents
Coding agents now work for hours. Over a run they have questions. They hit blockers. They produce plans and prototypes and diffs you need to review. The Inbox is where all of that arrives while they keep working. You read it like mail and answer when you get to it. Reviewing and annotating is the native Plannotator experience. The agent picks up your answer and goes on.
Local. Claude Code and Pi and OpenCode and any MCP agent. Open source.
Set up the Inbox
Install Plannotator
curl -fsSL https://plannotator.ai/install.sh | bash irm https://plannotator.ai/install.ps1 | iex Start the Inbox
plannotator inbox It opens in your browser. You only run this once: from then on, agents start the Inbox in the background when they write to it.
Connect your agent
claude plugin marketplace add backnotprop/plannotator claude plugin install plannotator@plannotator
Then quit Claude Code and start it again. The plugin adds a plannotator_inbox tool
when Claude Code starts, once the Inbox has run, and your replies arrive as Claude's next turn.
Claude Code 2.1.287 or newer on macOS or Linux. On Windows, use the MCP command under Other MCP agent.
pi install npm:@plannotator/pi-extension
Then turn on Pi under The Inbox in your agent's tools in the Inbox's Settings, and start a new Pi
session or run /reload. Your replies arrive as Pi's next message once it is idle.
The tool is off by default on Pi because Pi sends every tool's definition with each request.
{ "plugins": [{ "package": "@plannotator/opencode@latest" }] } { "plugin": ["@plannotator/opencode@latest"] } Then turn on OpenCode under The Inbox in your agent's tools in the Inbox's Settings, and restart OpenCode. On OpenCode 2 your replies arrive as the session's next turn.
codex mcp add plannotator-inbox -- plannotator inbox mcp Then restart Codex. Codex reads your reply when it calls wait_for_reply.
If it can't find plannotator, use the full path from command -v plannotator.
{ "mcpServers": { "plannotator-inbox": { "command": "plannotator", "args": ["inbox", "mcp"] } } } Cursor reads your reply when it calls wait_for_reply.
If it can't find plannotator, use the full path from command -v plannotator.
plannotator inbox mcp
Add it to your agent as a stdio MCP server named plannotator-inbox. The Inbox's first
screen gives the exact step for VS Code, Gemini CLI, Windsurf, Zed, Goose, Amp, Cline, Kiro,
JetBrains and the Claude app, with the full path to plannotator.
Send yourself a test message
Ask your agent, in a session started after step 3:
Send me a test question in the Plannotator Inbox, then wait for my answer. Pick an answer in the Inbox and press Send. The thread then reads "Delivered to" your agent.
Why not the agent's own window
The agent's window is for the run you are watching. Long runs are the ones you are not watching. With several agents going you are not sitting in any of their windows. Their questions and blockers and reviews need somewhere to land that is not a scrollback.
The Inbox is that place. Every agent writes to it. Each item is a thread with its context attached. You open it when you are ready. You answer or annotate once. The agent continues from there.
How it works
An agent writes to the Inbox while it works.
A question with options. A blocker. A plan or a prototype or a diff to review.
You answer in the Inbox.
Pick an option. Annotate the file with Plannotator. Send.
The agent gets the answer.
It arrives as the session's next turn. The thread shows "Delivered to Claude Code".
What arrives during one run
01
The plan
The agent drafts a plan and asks what it cannot decide alone. The plan opens next to the thread. You comment on a line and pick the answers. Nothing reaches the agent until you send.
If the agent edits the plan after sending it the Inbox says so and opens the version it sent.
You are somewhere else while this happens. The agents keep working. The Inbox tells you when something needs you.
02
A prototype
The agent attaches the prototype. It opens inside the Inbox at full screen. You pin comments on the parts to change. The comments go with your next send.
Markdown and HTML and Mermaid and Graphviz. Plannotator's annotation tools.
The prototype open in the Inbox with one comment pinned
03
The guided review
At the end of the run the agent sends a guided review. The change is split into sections in reading order. Each section has its diff and a short summary. Every file is checked against the real diff. You read the whole change in one sitting.
Connections
Plannotator's Claude Code mod and Pi extension and OpenCode plugin write to the Inbox on their own and wake the agent when you send. Claude Code has the Inbox tool on by default; for Pi and OpenCode, turn it on in the Inbox's Settings (it applies from the next Pi session or OpenCode start).
Claude Code
Pi
OpenCodeAny other agent connects to the Inbox as a local MCP server. The Inbox shows the exact step for Codex and Cursor and VS Code and Gemini CLI and Zed and Goose and Amp and Cline.
codex mcp add plannotator-inbox -- plannotator inbox mcp Local
- Listens on
127.0.0.1. Opens in your browser. - Threads and answers are files under
~/.plannotator. - Part of Plannotator. MIT or Apache 2.0.
coming The same inbox hosted and on your phone.
Blog An Async Surface for Long-Running Agents
Install
Install Plannotator and run plannotator inbox.
curl -fsSL https://plannotator.ai/install.sh | bash 
