Agent Inbox collects questions, approvals, plans, and status updates from coding agents. Answer in one place instead of hunting through your sessions.
3 open requests across your projects
The portal is ready for authentication. Using your existing provider keeps accounts in one place.
This choice determines how customers access the portal. After your decision, the agent will wire up sign-in and add the integration tests.
Waiting for agent pickup.
Agent Inbox turns scattered terminal interruptions into a shared, durable workflow for the decisions only you can make.
Based on maintainer usage.
Bring decisions and human tasks out of individual agent sessions.
Keep the action, tradeoffs, background, and next step together.
See what waits on you, what the agent picked up, and what finished.
Keep progress, responses, and outcomes attached to stable work.
Open a request to see the project and what needs your attention. Respond there, with the background close by.
Each question includes the decision and why it matters. Choose an option, write your own answer, ask for clarification, or snooze it. Open approvals can also show a Preview link for the current PR head when your project publishes successful, non-production GitHub deployment metadata.
Try a sample questionWrite to a connected session from its panel in the Agents view. Before you send, Agent Inbox says how the message will arrive. Claude Code sessions with the wake-up hooks can wake for it, and so can a Copilot CLI session that left its message watcher running. A Claude Code session started with the channels flag gets it pushed straight in; channels are a Claude Code research preview that needs a claude.ai or Console login. Other sessions see it when they next check Agent Inbox. The agent acknowledges with an outcome, and messages to sessions that ended are listed under Not delivered.
Plans keep tasks, responses, and recorded outcomes together, so you can see what happened on each project. An agent can link a plan to a GitHub issue. The plan then lists the issue's sub-issues and how many are done, and rows linked to a sub-issue show their parent. Agent Inbox looks these up through your signed-in GitHub CLI.
The Agents view gives each session a card with its latest report, its plan's progress, its activity over the last 30 minutes, and anything it is waiting on you for. Open a card to answer, snooze, or pin a request in its panel; a request from a session that has ended gets its own card, so nothing waits out of sight. With Agent Inbox's Claude Code or Copilot CLI hooks installed, a session busy with other tools shows as working even before it reports. Background refreshes keep it current without pulling keyboard focus away from chat or another app; in the active viewer, keyboard focus stays where you left it.
In the macOS app, choose New session to start a Copilot CLI session in one of your projects. Agent Inbox shows what will start before it does. Each tool call that needs approval arrives in the inbox as Allow once or Deny, and you can message, park, resume, or stop the session from its card in the Agents view. Each session works in its own git worktree, so several can run in one project at once. Edits in the worktree lets a session write inside its folder without asking; everything else still asks. On a source checkout, a terminal driver does the same with approvals answered in that terminal. The launcher is off until you turn it on in Settings.
Agents send requests through MCP. You answer in Agent Inbox, and they read your response through the same connection.
Connected through MCPIt sends the question and background to Agent Inbox, along with the project it belongs to.
Pick an option, write your own answer, or mark a task done. You can also ask for clarification or snooze the request.
Your agent checks MCP for responses. You'll see when it has picked up your answer and when it records an outcome.
Questions, plans, and responses are stored in a local SQLite database. Agent Inbox doesn't call a model or run a hosted service.
Start with the desktop app, or run it from a Node.js checkout.
After installing, open Settings → Setup and choose your coding tools. Start a fresh CLI session to load the connection.
In the macOS app and the Linux AppImage, Settings → Updates can download a new release, check it against the signed release manifest, and install it when you choose Install and restart. Setup's Update agents button then moves your agents to the new version.
Get the desktop app Installation guide macOS 13.5+ on Apple silicon · Linux x64 & arm64Use Node.js 24. These commands build the app and preview the agent setup.
git clone https://github.com/shariqh/agent-inbox.git
cd agent-inbox
npm ci
npm run build
npm run install:agentsNo. It stores the questions, plans, and updates your coding agents send, along with your replies. It doesn't write code or make model calls.
Any app that can connect to a stdio MCP server can use Agent Inbox. The installer automates setup for GitHub Copilot CLI and Claude Code; other MCP clients need manual configuration.
Your inbox is stored locally, with no Agent Inbox telemetry or hosted service. Optional update checks and downloads contact GitHub, and so do pull request and issue lookups when the GitHub CLI is installed and signed in. Your coding tools still have their own data and model policies.
Agent Inbox shows Preview only on an open approval question or blocked approval plan row when GitHub reports a successful, explicitly non-production deployment for the current PR head. Your project must publish standard GitHub deployment metadata. Missing, stale, or conflicting records leave the Preview link out; the approval controls stay available.
Agents read replies through the MCP pending tool. Copilot CLI can use background watchers for questions and blocked plan actions; Claude Code has opt-in wake-up hooks. Wake-up depends on your setup. The receipt shows whether the agent has picked up your answer. On a Copilot CLI request, choose "I'll take a while" to tell the agent to stop waiting, or to wait quietly for a few hours or until tomorrow. If it stops, prompt it in its terminal after you answer. The request stays in your inbox. Agents read a message you send a running session through the same pending tool. A Copilot CLI session sees it when it next checks Agent Inbox, or at once if it left its message watcher running, and a Claude Code session with the hooks can wake for it. A Claude Code session started with the channels flag, a research preview that needs a claude.ai or Console login, gets messages pushed straight in.
Desktop packages are available for macOS 13.5+ on Apple silicon, and Linux on x64 and arm64. Intel Macs stay on v1.8.0, the last version for them. You can also run the local browser viewer from a Node.js 24 checkout. See the installation guide for Linux runtime requirements and package verification. Windows and hosted deployments are not currently supported.