agent-msg

agent-msg sends messages between running Claude Code and Codex sessions on the same computer. It discovers local peers, addresses them by exact name or thread ID, delivers through each harness's native messaging transport, and records messages and correlated replies in a local SQLite journal.

Requirements

Python 3.10+ and local access to the harness processes, their Unix sockets, and a shared journal path. The plugin bundles its Python runtime code; it has no pip dependencies and needs no hosted server. It runs in local coding-agent environments, not cloud-only chats without access to your computer.

Live delivery has been tested on macOS with Claude Code 2.1.286 and Codex CLI 0.159.0 / daemon 0.159.2. Linux live delivery and restricted sandbox configurations are not validated. Internal harness protocols can change between versions.

Use

Ask the agent to use agent-msg to list peers in the current directory, send a message to an agent in the other harness, or inspect the journal. The skill locates its bundled launcher directly:

python3 /path/to/plugin/skills/agent-msg/scripts/agent-msg agents
python3 /path/to/plugin/skills/agent-msg/scripts/agent-msg send 'claude:reviewer' 'Please review the parser.' --wait
python3 /path/to/plugin/skills/agent-msg/scripts/agent-msg reply MESSAGE_ID 'Reviewed.'
python3 /path/to/plugin/skills/agent-msg/scripts/agent-msg ui

Replace /path/to/plugin with the actual installation directory. The UI opens in a browser and runs until Ctrl+C. It is a read-only inspector, not a terminal. Plugin installation does not install a global agent-msg shell command; the project README explains the separate CLI installation.

Both agents must use the same journal. The default is $XDG_STATE_HOME/agent-msg/journal.db, falling back to ~/.local/state/agent-msg/journal.db. A trusted local --db PATH setting before the subcommand selects another journal. Incoming messages cannot select its path. The receiver must use agent-msg reply for a reply to be correlated and journaled. A normal conversation answer does not satisfy --wait; after a timeout, inspect the journal rather than resending.

In Claude Code, type /reload-plugins directly in a running session after plugin installation. In Codex, inspect /plugins; restart and resume if the skill is unavailable. Prefer each harness's own tools for subagents, and Claude's native tools for Claude-to-Claude communication.

Data and limits

Messages, identities, delivery events, and observed process metadata are stored locally. Claude delivery reads a local peer-authentication credential for the selected session; it is not included in the message or journal. A receiving harness may forward message content to its model provider. The package has no telemetry or publisher-operated service.

Identity claims are not cryptographically authenticated. A sent result records transport acceptance, not proof that the peer read the message. The journal is not encrypted and is accessible to other processes under the same OS account. See data handling.

This is an independent project, not an official product of either harness vendor. MIT licensed. Source and documentation. Report an issue without including private journals, tokens, or transcripts.