agent-inbox

Implements file-based messaging between AI coding agents using markdown files and atomic filesystem operations.

1|Updated Nov 11, 2012
One-click install
npx skills add https://github.com/fairchild/dotfiles --skill agent-inbox-fairchild
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: agent-inbox
Source: https://github.com/fairchild/dotfiles/tree/main/agents/shared/first-party-skills/agent-inbox
Command: npx skills add https://github.com/fairchild/dotfiles --skill agent-inbox-fairchild

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve? AI coding agents running in different tools (Claude Code, Codex, Cursor, Gemini CLI, Warp) have no built-in way to coordinate on the same project. This Skill provides a daemon-free, network-free messaging protocol where agents exchange markdown messages through a shared filesystem inbox. ## Core Features & Use Cases - Cross-harness messaging: Send and receive messages between agents in any tool using only mkdir, cat, and mv — no MCP server, daemon, or registry required. - Atomic delivery: Messages are written to tmp/ then moved to new/, preventing partial reads; read messages are archived for audit. - Notification hooks: Optional Stop and SessionStart hooks surface unread mail with a one-line summary, plus a cmux-based wake script that can spawn a headless session to process replies. - Use Case: Two agents — one rewriting auth middleware, one updating the frontend — coordinate by dropping timestamped markdown messages into each other's inbox directories inside the repo-shared .agents/inbox root, with reply_to paths enabling replies without prior setup. ## Quick Start Ask the agent to set up an agent inbox for this repository and check for any unread messages.

Frequently Asked Questions about agent-inbox

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I send messages between AI coding agents?▼

Write a markdown file with YAML frontmatter (from, to, reply_to, timestamp) into the recipient's tmp/ directory, then move it to their new/ directory. The recipient reads it with cat and archives it with mv — no server or daemon needed.

How do agents in different tools like Claude Code and Cursor coordinate?▼

They share a filesystem inbox rooted beside the repo's common git directory, so all worktrees and tools see the same .agents/inbox tree. Discovery is human-brokered: you tell each agent the other's inbox path, and every message carries a reply_to path for responses.

Does agent-inbox require a daemon, MCP server, or network access?▼

No. The protocol uses only mkdir, cat, and mv on the local filesystem. Messages are markdown files written atomically via tmp/ then renamed into new/, making it work entirely offline.

How do I get notified of unread agent messages?▼

Two optional hooks handle notification: a Stop hook (check-inbox-hook.sh) prints a one-line unread count after each turn, and a SessionStart hook (inbox-startup.sh) summarizes unread messages with sender and subject when a session begins. Both are silent when the inbox is empty.

What are the limitations of file-based agent messaging?▼

Delivery is pull-based — recipients only see messages when they check their inbox or a hook fires, so there is no real-time push. Discovery is manual, and the wake-parent script depends on cmux and uses --dangerously-skip-permissions for headless sessions.