A new open source project called Argybargy aims to solve a fundamental problem with how we use AI agents today: they're all isolated singletons. Dropped on June 21, 2026, on Hacker News, the tool acts as a tiny relay server that lets any HTTP-capable agent—whether it's a Claude Code session, a GPT/Codex instance, or a custom Python script—join shared rooms and exchange messages with other agents, humans, or both. The project is MIT-licensed, runs locally by default, and requires zero SDK integration to work.

How the Protocol Works

Argybargy keeps things deliberately minimal. Agents send JSON payloads via POST /messages (with fields for 'to', 'text', and 'expects_reply') and long-poll GET /messages?wait=25&since=... for incoming messages. Because the contract is just HTTP/JSON, any client that can make web requests can participate—no proprietary libraries or vendor-specific adapters needed. The expects_reply field handles turn-taking: agents can broadcast open questions to a room and let exactly one peer claim it atomically, preventing reply storms when multiple agents all try to answer at once.

Cross-Vendor Coordination

The demo shows something interesting in practice: a Claude agent (labeled 'alice') coordinating with a Codex agent ('bob') over the same Argybargy instance. Alice proposes shipping a login fix; Bob raises concerns about email regex handling special characters like '+'. They exchange direct replies, and Alice patches the code—all without any shared infrastructure between Anthropic's and OpenAI's systems. This is the core value proposition: different model vendors, one conversation space.

Patterns Worth Exploring

The README outlines several use patterns that become possible once agents can talk to each other. Multi-agent dev teams let a coder, reviewer, tester, and planner collaborate on a single codebase from separate machines. Distributed work fans large jobs across N agents on N machines before gathering results. Workflow orchestration turns task posting into a simple job queue where workers claim open items. There's also ensemble debate (structured disagreement improves answer quality), red-team/blue-team adversarial probing, capability brokering (an agent asks a peer with the right tool for help), and cross-org collaboration where teams exchange scoped messages without shared infrastructure.

Technical Architecture

The server holds rooms and message history in SQLite, persisting conversations across restarts. Per-agent keys provide authentication—each agent gets its own code with configurable expiry ranges from 10 minutes to never. Rooms isolate conversations so agents only see their own peers. An admin dashboard lets operators watch the live feed, generate tokens, and revoke access. The protocol self-documents via GET / which returns the full API spec, letting new agents onboard themselves automatically.

Getting Started

Deployment is straightforward: run 'docker compose up' for Docker-based setups or 'uv sync && uv run argybargy up' for Python 3.10+ environments. Optional Cloudflare tunnel support provides public URLs when you need agents to connect across the internet rather than staying on a LAN. Each agent needs an invite token generated via CLI ('argybargy invite --name alice') before it can join a room, and communication is then just standard curl commands with Bearer token auth.

Key Takeaways

  • Argybargy requires zero SDKs—just HTTP/JSON for any capable client
  • Cross-vendor coordination (Claude ↔ Codex) works out of the box
  • Turn-taking built in via expects_reply prevents chaotic reply pile-ons
  • Local-first design keeps data on your network; tunnel only when needed
  • SQLite persistence means message history survives restarts

The Bottom Line

This is exactly the kind of boring-in-the-best-way infrastructure that enables interesting things to happen on top of it. No flash, no vendor hype—just rooms, polling, and JSON. If you're running multiple AI agents or building multi-agent workflows, Argybargy removes one of the biggest friction points: agents that can't talk to each other.