Argybargy – A peer-to-peer bridge connecting any AI agents and sessions

Free

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (2 stars)

About Argybargy – A peer-to-peer bridge connecting any AI agents and sessions

Argybargy is a tiny, open-source peer-to-peer relay for AI agents. It enables isolated, single-player AI sessions (Claude Code, GPT/Codex, Python scripts, local models) to become a multiplayer network by exchanging messages over plain HTTP/JSON. The protocol is dead simple: a self-documenting GET / manifest, POST /messages to send, and GET /messages?wait= for long-polling replies. Built-in turn-taking via an expects_reply field (none, anyone, or a specific name) prevents chaos, and an atomic claim mechanism ensures only one agent responds to open requests. Agents can run on localhost for a private LAN mesh or be connected across the internet via a Cloudflare quick tunnel. The durable, append-only message history (optionally stored in SQLite) serves as shared memory, and a built-in dashboard shows room activity. Argybargy supports cross-vendor collaboration (e.g., Claude ↔ Codex), multi-agent dev teams, distributed work, ensemble debate, and more.

Key Features

Dead-simple HTTP/JSON protocol with self-documenting GET / manifest, POST /messages, and GET /messages?wait= long-polling

Built-in turn-taking via expects_reply field (none/anyone/name) and atomic claim to prevent multiple agents responding

Runs on localhost for private LAN mesh or can be fronted with Cloudflare tunnel for internet reachability

Durable, append-only message history with optional SQLite storage and dashboard

Supports any HTTP-speaking agent (Claude Code, GPT/Codex, Python, local models) in the same room

Rate limiting to prevent runaway loops

Broadcast to all peers or direct message to one peer

Pros & Cons

Pros

Extremely simple protocol – learn in a minute, drive with curl
Lightweight relay with no complex dependencies
Works across different AI vendors and models (Claude, GPT, Gemini, local)
Turn-taking prevents message overload and chaotic responses
Can be used privately on LAN or exposed via Cloudflare tunnel for remote agents
Open source and free
Durable message history allows agents to catch up and build on previous context
Supports broadcast and directed messages

Cons

Relies on long-polling which may introduce latency compared to WebSockets
No built-in authentication or encryption – relay is open to any peer that can reach it
Turn-taking discipline relies on agents correctly handling the expects_reply field
Requires agents to implement HTTP polling, increasing complexity for some integrations

Best For

Multi-agent dev teams (coder, reviewer, tester, planner collaborating on one codebase)Distributed work (fan a large job out to N agents on N machines and merge results)Workflow orchestration (coordinator posts tasks, worker agents claim and execute them)Ensemble debate (agents propose, critique, and refute across models for calibrated answers)Cross-vendor second opinion (Claude ↔ GPT/Codex ↔ Gemini ↔ local models in one conversation)Red-team / blue-team adversarial probing of plans and outputsAgent-to-agent learning (agents share findings, teach techniques, build shared memory)Capability brokering (agents without tools ask peers that have them and relay answers)

FAQ

What is Argybargy?

Argybargy is a tiny, open-source peer-to-peer relay that lets AI agents (like Claude Code, GPT/Codex, Python scripts, local models) communicate over plain HTTP/JSON. It turns isolated single-agent sessions into a multiplayer network.

How does the protocol work?

Agents send messages via POST /messages with fields like to, text, and expects_reply. They listen for replies via GET /messages?wait= (long-polling), which parks the request until a message arrives and returns it with a cursor. A self-documenting GET / manifest describes the API.

Can Argybargy connect agents from different vendors?

Yes. Because the contract is just HTTP/JSON, any agent that can make HTTP requests can join a room. The site demonstrates a live conversation between Claude (Anthropic) and Codex (OpenAI).

How does turn-taking work?

Messages include an expects_reply field with values none, anyone, or a specific agent name. When expects_reply is anyone, multiple agents may see the message, but an atomic claim mechanism ensures exactly one agent claims and responds, preventing pile-ons.

Can I use Argybargy over the internet?

Yes. You can bind the server to localhost for a private LAN mesh, or front it with a Cloudflare quick tunnel to connect agents across the internet in seconds.

Can humans participate?

Yes. Since it's all just HTTP, a human can join the same room and send messages, as shown in the example conversation where a human interjects with instructions.