agmsg is a small transport layer for teams that run more than one CLI coding agent at a time. Instead of asking Claude Code, Codex, Gemini CLI, Copilot CLI, or OpenCode to hand work back through a human, each agent writes to and reads from a shared local SQLite room. The project is deliberately plain Bash plus sqlite3: no daemon, no hosted broker, no MCP server, and no framework lock-in. That makes it a useful fit for multi-agent coding experiments where the main problem is coordinating independent terminal sessions rather than building another orchestration runtime.
The current README documents installs through npx/npm, a setup.sh one-liner, git clone, and the Claude Code plugin marketplace. Once installed, agents can use slash-command style entry points such as /agmsg or $agmsg, pick a delivery mode, join a team room, send messages, monitor incoming work, and replay history from the durable database. The author explicitly distinguishes agmsg from subagents: spawn can open a peer agent in another terminal, but the relationship remains peer-to-peer messaging over the local store rather than parent-child task management.
The main caveat is scope. agmsg does not replace MCP tools, agent frameworks, review queues, or enterprise collaboration policy; it gives compatible local agents a simple message floor. Teams should still decide how prompts, repository permissions, secrets, logs, and provider accounts are governed across every participating agent. Its Product Hunt and GitHub launch traction make it a timely page for aicoolies, but the buyer framing should stay grounded in local coordination, transparent SQLite history, and low operational complexity rather than claiming production-grade orchestration.
