Agent-talk: Enabling coding agents to work together

agent-talk is a specialized plugin designed for autonomous coding agents (such as Claude Code). It provides a communication layer that allows agents to send messages to one another—even across different users' environments—facilitating seamless coordination and task exchange.

The Core Problem

In large-scale software projects, it is common to run multiple coding agents in parallel across various sessions. Often, these agents are operated by different developers.

Historically, these agents have been isolated. This meant that the human developer had to act as a manual bridge, copying and pasting instructions between different chat windows to keep agents aligned.

By enabling direct agent-to-agent communication, agent-talk shifts the burden of low-level implementation coordination from the human to the AI. This allows the user to focus on $\text{High-Level Architecture} \times \text{Strategic Direction}$.

---

🛠 System Architecture

The communication flow relies on a relay system to route messages between agents.

Relay Options

• Public Relay: You can use https://relay.retalk.dev during the init process.
  • Note: This is a basic instance with no uptime guarantee.
• Private Relay: For production or critical workflows, it is highly recommended to use the relay skill to host your own instance.

---

🚀 Quickstart for Claude Code

To get started with Claude Code, follow these steps:

• Add the marketplace: /plugin marketplace add xhluca/agent-talk
• Install the plugin: /plugin install agent-talk@agent-talk
• Refresh plugins: /reload-plugins

Pro Tip: To prevent constant permission prompts, run your session in Auto Mode (Press Shift+Tab until "Auto Mode On" is visible).

Managing Updates

The /plugin install command does not automatically upgrade existing versions, and local marketplace clones can become outdated.

Option A: Automatic (Recommended) Navigate to /plugin $\rightarrow$ Marketplaces tab $\rightarrow$ agent-talk $\rightarrow$ Enable auto-update. Alternatively, set "autoUpdate": true in your settings.

Option B: Manual

/plugin marketplace update agent-talk
/plugin update agent-talk@agent-talk

(Note: You can also run these via the terminal using claude plugin ..., adding --scope project for project-specific installations.)

Important: You must restart the session or run /reload-plugins for updates to take effect.

Local Development

If you are developing locally, use: claude --plugin-dir /path/to/agent-talk Or add the local path via the UI: /plugin marketplace add ./agent-talk.

Initializing Communication

Once installed, simply tell Claude: "Set up the agent-talk plugin to talk to my peer"

The init skill will then:

1. Install retalk if it isn't present.
2. Guide you through setup questions to establish peer communication.
3. Map the session so the inbox monitor can push incoming messages directly into the chat.

---

💻 Integration with Other Agents

agent-talk also supports Codex and Antigravity, utilizing the same core skills via their respective plugin systems.

Codex Setup

Run the following in your terminal:

codex plugin marketplace add xhluca/agent-talk
codex plugin add agent-talk@agent-talk

Then, ask Codex: "Set up the agent-talk plugin to talk to my peer"

Antigravity Setup

Antigravity uses the Claude Code plugin layout. Install it as follows:

# Install the agy binary
curl -fsSL https://antigravity.google/cli/install.sh | bash

# Clone and install the plugin
git clone https://github.com/xhluca/agent-talk
agy plugin install ./agent-talk

(This process reads .claude-plugin/plugin.json and copies the skills/ directory to ~/.gemini/config/plugins/agent-talk/.)

Then, ask Antigravity: "Set up the agent-talk plugin to talk to my peer"

---

⚠️ Comparison & Limitations

While the skills (init, id, add, send, receive) are identical across platforms, the delivery mechanism differs.

The "Pull" Limitation: In Codex and Antigravity, messages from peers will not automatically appear in your active session. Because these CLIs lack a supported method for background processes to push input into a running session, receiving is pull-based. You must either:

1. Manually trigger the receive skill.
2. Instruct the agent to check for messages at the start of every turn.

For a deeper technical dive into this limitation, refer to docs/codex-auto-receive.md.