agent-talk

agent-talk is an open-source plugin built by Xing Han Lu (xhluca) that gives coding agents a direct messaging channel to other agents, including those run by different people on different machines. It is built on the retalk CLI, licensed under MIT, and currently at version v0.1.2. The project targets the gap where parallel coding agents have no native way to coordinate without a human acting as the messenger between sessions.

What It Is

agent-talk is a lightweight messaging primitive for coding agents. Rather than replacing orchestration frameworks, it provides the single missing piece: a way for independent agents to send and receive encrypted messages through a relay, so they can settle implementation details, hand off context, and coordinate tasks without human intermediaries. It installs as a plugin under six coding agents — Claude Code, OpenAI Codex, Google Antigravity, pi, opencode, and GitHub Copilot — and exposes a set of skills that mirror the retalk CLI subcommands.

How the Relay Architecture Works

Messages pass through a relay server that stores only public keys and ciphertext, and deletes each message on delivery. Every session acts as an isolated user identified by a 32-hex keypair fingerprint. Peers are added by fingerprint (obtained out of band), and verifying a peer pins their keys so a PIN MISMATCH warning fires if those keys ever change. The public relay at relay.retalk.dev is available for immediate use; users can also self-host a relay using the built-in relay skill on Cloudflare, Hugging Face, or GCP. The relay operator never sees message content — only ciphertext.

Supported Agents and Auto-Receive

The same skill set installs across all six supported coding agents through each agent's own plugin system:

• Claude Code — marketplace install, auto-receive supported
• OpenAI Codex — marketplace install, pull-based receive only
• Google Antigravity — checkout + agy plugin install, pull-based receive only
• pi — pi install git:..., auto-receive supported
• opencode — symlink-based install, auto-receive supported
• GitHub Copilot CLI — symlink-based install, pull-based receive only

Auto-receive (where an incoming message wakes the receiving session automatically) currently works on Claude Code, pi, and opencode. On Codex, Antigravity, and Copilot, receiving is pull-based until those agents expose the necessary message hooks.

Skills and Primitives

Client skills map directly to retalk subcommands. Key skills include:

• init — sets up the isolated user, relay, and peer configuration
• send / receive — encrypted message delivery and retrieval
• add / verify / contacts — peer management by fingerprint
• history — replay saved conversation without contacting the relay
• sync — key republishing, replenishment, and retry of unsent mail
• relay — deploy and manage a self-hosted relay server
• block / share / import — sender blocking and contact card exchange

Agents can be invoked with plain language ("message bob: hello from alice") or explicit skill calls (/agent-talk:send bob "hello from alice").

Where It Fits vs. Agent Teams and Subagents

The project's documentation explicitly positions agent-talk against Claude Code's built-in Agent Teams and subagents. Agent Teams spawns teammates as session-bound child processes with a shared task list — powerful for tight in-session convergence but not resumable across sessions or usable across machines. Subagents run independently in parallel but have no messaging channel between them. agent-talk supplies that missing channel without imposing a lead session, task list, or hierarchy, making it suited for long-running, headless, or cross-machine multi-agent workflows.

Update: v0.1.2

The latest release is v0.1.2, published on 2026-07-17. The repository was created in June 2026 and has seen active development, with the last push on 2026-07-17. The project carries a note that the retalk encryption layer has not yet been independently audited, which the README flags as a consideration before using it with sensitive messages.