fujibee/agmsg

What it solves

agmsg eliminates the need for users to manually copy and paste messages between different CLI AI agents. It allows agents from different tools (such as Claude Code, Codex, Gemini CLI, and GitHub Copilot CLI) to communicate directly with one another through a shared local database.

How it works

The project acts as a thin transport layer using a shared SQLite file in WAL-mode. Each agent is equipped with a hook or a monitor stream that reads incoming messages from this database and surfaces them as text the agent can react to. Sending a message is handled by a simple script that appends a row to the SQLite file. There is no central daemon, network socket, or broker required.

Who it’s for

Developers using multiple CLI-based AI agents across different projects who want these agents to collaborate, perform peer reviews, or coordinate tasks without human intervention.

Highlights

• Cross-Agent Compatibility: Supports a wide range of agents including Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Antigravity, and OpenCode.
• Zero Infrastructure: Requires only bash and sqlite3, avoiding the complexity of MCP servers or external runtimes.
• Agent Spawning: Ability to spawn new agent processes in separate terminal windows or tmux panes and assign them specific roles.
• Flexible Delivery Modes: Offers multiple delivery options, including monitor for real-time push and turn for checking messages between assistant turns.
• Role Management: Supports switching identities within a project using actas and drop commands.
• Session Persistence: Integrates with tmux-resurrect to bring back agent roles with their previous conversation context.