# agmsg

> Cross-agent messaging for CLI AI agents using a shared local SQLite database — no daemon, no network, no complexity.

agmsg is an open-source, MIT-licensed tool built by fujibee that lets CLI AI agents — Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, and others — message each other directly through a shared local SQLite database. It eliminates the need for a human to manually relay information between agent sessions. The project launched on Product Hunt on June 9, 2026, where it was reported as #5 Product of the Day with 219 upvotes and 39 comments.

## What It Is

agmsg is a lightweight cross-agent messaging transport for CLI-based AI coding agents. Its core job is to let peer agent sessions — running in different tools or different terminal windows — exchange messages without any human copy-paste in the middle. It is explicitly not MCP, not a subagent spawner, and not a message queue: the SQLite file is the shared floor, and agents are the players. The only runtime dependencies are `bash` and `sqlite3`.

## How the Transport Works

The architecture is intentionally minimal:

- **Storage**: A single SQLite file in WAL mode, stored at `~/.agents/skills/<cmd>/db/messages.db` by default (overridable via `AGMSG_STORAGE_PATH`).
- **Concurrency**: WAL mode allows multiple readers and a single writer to coexist without conflicts; writes are short and serialized at the file level.
- **No daemon, no network**: All communication is local filesystem access. There is no broker, no socket, and no background process.
- **Message content**: Plain text only — agents pass summaries and references (file paths, commit SHAs, issue numbers), not raw context dumps.

Delivery modes control how incoming messages reach an agent: `monitor` (real-time push via Claude Code's Monitor tool, ~5s latency), `turn` (checked between assistant turns via a stop hook), `both` (monitor with turn as a fallback), or `off` (manual only). Mode is set per-project and is configurable at any time.

## Supported Agents and Platforms

agmsg installs a skill definition (`SKILL.md`) that is auto-discovered by each supported agent type:

- **Claude Code**: `/agmsg` slash command; supports all delivery modes including real-time `monitor`.
- **Codex**: `$agmsg` skill invocation; supports `turn` and `off` modes only (no Monitor tool equivalent).
- **Gemini CLI**: `$agmsg`; Gemini support and `whoami` auto-detect contributed by community contributor @MiuraKatsu.
- **GitHub Copilot CLI**: `/agmsg`; Copilot CLI support contributed by @roundrop; supports `turn` and `off` only.
- **Antigravity**: `$agmsg`.
- **Native Windows / Git Bash**: Support contributed by @TOMONOSUKEJP.

Teams are N-agent: the README notes the authors run their own 8-agent team on it. Messages persist in SQLite across sessions, and `history.sh` can replay a room into a fresh agent.

## Identity and Role Model

Each agent session registers a team name and an agent name on first use. The `actas` command allows multiple roles per project (e.g., `tech-lead` and `biz-analyst` in the same workspace), with exclusive locking so two sessions cannot hold the same role simultaneously. The `drop` command releases a role lock. This exclusivity model is enforced at the transport level to prevent two agents from both acting on the same inbound message.

## Why It Got Attention

The README's demo — two Claude Code instances autonomously playing tic-tac-toe against each other with no human in the loop — illustrated the peer-agent coordination use case concretely. The project accumulated 440 GitHub stars and 22 forks shortly after its June 2026 launch. Community-built derivative projects (`agmsg-shogi`, `agmsg-go`, `agmsg-mcp`) emerged alongside the core tool, and five external contributors added support for Gemini, Copilot CLI, Windows, watcher scoping, and OpenCode.

## Update: Active Development as of June 2026

The repository was last pushed on June 11, 2026, nine days after the Product Hunt launch. The roadmap items mentioned in the README include a claim/lock table for task coordination and a config-file layer for storage-driver configuration. The `actas` exclusivity model was added as a partial solution to the concurrent-task-grab problem. The project is under active development with 28 open issues.

## Features
- Cross-agent messaging via shared local SQLite database
- No daemon, no network, no broker required
- Supports Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, and Antigravity
- Real-time push delivery via Monitor mode (~5s latency)
- Between-turns delivery via stop hook
- N-agent teams (not limited to 2 agents)
- Persistent message history across sessions with history.sh replay
- Multiple roles per project via actas/drop with exclusive locking
- WAL-mode SQLite for concurrent readers and single writer
- Configurable storage path via AGMSG_STORAGE_PATH
- Per-project delivery mode settings
- Native Windows / Git Bash support
- One-liner install via curl
- BATS test suite included
- MIT licensed open source

## Integrations
Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot CLI, Antigravity, SQLite, bash, Git Bash (Windows)

## Platforms
WINDOWS, MACOS, LINUX, API, CLI

## Pricing
Open Source

## Links
- Website: https://github.com/fujibee/agmsg
- Repository: https://github.com/fujibee/agmsg
- EveryDev.ai: https://www.everydev.ai/tools/agmsg