fujibee/agmsg: Cross-vendor messaging for CLI AI coding agents

What changed

The fujibee/agmsg project provides a direct messaging layer for command-line AI coding agents, enabling them to communicate with each other without human intervention. This system supports cross-vendor compatibility, allowing agents such as Claude Code, Codex, Gemini CLI, and GitHub Copilot CLI to form collaborative teams. The core mechanism relies on a shared local SQLite database, acting as a communication floor where agents can post and retrieve messages. Unlike traditional messaging systems, agmsg eschews daemons, network servers, or complex frameworks, instead utilizing only bash and sqlite3. This design choice simplifies setup and reduces potential points of failure.

The system operates by having each agent hook into a shared SQLite file. Incoming messages are surfaced as text that the agent can process, while outgoing messages are appended as new rows to the database via a send.sh call. The SQLite database is configured in WAL-mode, which allows multiple readers and a single writer to operate concurrently without conflicts. Message history is persistent, remaining in the database even after a session ends, and can be replayed into new agents using history.sh.

Installation can be done via npx agmsg for a quick, one-shot setup, or through npm i -g agmsg && agmsg install. Alternatively, developers can clone the repository and use the install.sh script for more customization, including setting a custom command name or agent type. For Claude Code users, a plugin marketplace option is available. The project emphasizes that different installation paths might offer varying levels of recency, with direct script clones always tracking the main branch, while tagged releases for npm packages or plugins may lag slightly.

On first use, agmsg prompts for a team name and an agent name for the current project. After this initial setup, users can interact with their agents using natural language commands like "send alice a message saying the deploy is done" or "check my messages." The system also supports advanced features such as actas for assigning multiple roles to a single project session (e.g., tech-lead and biz-analyst) and spawn for launching entirely new, independent agent processes.

Why it matters for builders

For AI builders, agmsg offers a foundational tool for creating more cohesive and capable AI agent teams. By abstracting away the complexities of inter-agent communication, developers can focus on designing agent behaviors and workflows rather than managing message passing. This enables the construction of more sophisticated AI systems where different specialized agents can collaborate on complex tasks, such as code review, debugging, or even collaborative development, all orchestrated through a unified messaging layer.

Practical impact

Developers can integrate agmsg into their existing CLI agent workflows to enable seamless communication between different AI coding assistants. For instance, one agent could be tasked with generating code, while another reviews it, and a third handles deployment checks, all communicating their status and findings through agmsg. The actas feature allows a single project to have multiple distinct AI personas, enabling nuanced task delegation and role-specific analysis. Builders should consider experimenting with npx agmsg for quick testing or cloning the repository for deeper integration and customization.

Caveats and source limits

The source material indicates that while agmsg supports various CLI agents, its implementation is primarily Bash-script based, with specific considerations for Windows environments requiring Git Bash. There is no mention of native support for PowerShell or other scripting languages on Windows. The project is described as a "fresh release" with "4 AI signals, 7 developer signals," suggesting it is relatively new and community adoption metrics like stars (1021) and forks (91) are provided, but extensive independent benchmarks or real-world case studies are not detailed in the provided excerpts. The exact performance characteristics and scalability under heavy load are not specified.