Agent-talk lets coding agents message each other

Copying instructions between agent windows by hand is exactly the problem agent-talk is trying to remove. The open source plugin, posted on GitHub by xhluca, gives coding agents a way to message one another directly so they can coordinate work across sessions and even across different users.

It is built on the retalk CLI and is aimed at teams running multiple agents in parallel on larger projects. The pitch is simple: instead of a human acting as the relay between agents, the agents can exchange messages themselves while people stay focused on higher-level decisions.

The project currently supports six coding agents: Claude Code, OpenAI Codex, Google Antigravity, pi, opencode, and GitHub Copilot. The same core skills are available across them, including:

Recommended reading

Avito eyes free dating inside its app

• init
• id
• add
• send
• receive
• history
• sync
• relay

Setup varies by client, but the Claude Code quickstart is straightforward:

• /plugin marketplace add xhluca/agent-talk
• /plugin install agent-talk@agent-talk
• /reload-plugins

The plugin also needs a retalk relay URL. Users without one can point setup at the public relay at https://relay.retalk.dev, though the repository says that instance has no uptime guarantee and recommends creating a dedicated relay for anything important.

There are some platform limits. Auto-receive — where a peer message appears in an active session automatically — is available on Claude Code, pi, and opencode. On Codex, Antigravity, and the interactive Copilot CLI, receiving is still pull-based because those clients do not support a background process injecting input into a running session.

The repository’s example is practical: one agent managing a dataset can answer another agent’s question about data splits, preventing training leakage without waiting for two humans to trade messages in Slack.

One caveat from the project itself: while retalk is designed for end-to-end encryption, the code has not been independently audited yet, so the author warns against trusting it with sensitive messages.