Show & tell: agent-chat: peer-to-peer messaging for independent Pi agents (no orchestrator)

[Hysilens-Helektra]

https://github.com/Hysilens-Helektra/agent-chat
I built a small extension for a use case that kept biting me: running several independent Pi sessions (often in different worktrees) that share Docker containers, ports, databases, or other machine-level resources.

Git isolates the files, but the agents have no idea the others exist. One agent starts docker compose up, another stops/rebuilds the same stack, and everything collapses into manual intervention.

agent-chat is intentionally not an orchestrator.

• No primary agent, no subagents, no task graph, no file reservations, no crew/plan/work/review workflow.
• Just a lightweight peer-to-peer communication layer so independently running Pi instances can discover each other and exchange messages.
• Agents decide for themselves when to talk, what to say, and how to resolve conflicts.
• You can start (or stop) agents at any time; they join and leave dynamically.

How it works

• Tiny local TCP server (server.mjs) that binds only to 127.0.0.1. Zero dependencies.
• Extension (agent-chat.ts) registers three tools: agent_chat_register, agent_chat_list, agent_chat_send.
• Messages are delivered as ordinary user / steering messages. Delivery is push-based; the sender gets an ack status (delivered / buffered / failed) within a short window.
• Optional SKILL.md gives the agents guidance on when to communicate (resource contention, anomalies, etc.) without forcing a rigid protocol.

The whole thing is two source files + a skill. No extra runtime dependencies beyond Pi itself.

Comparison with related tools

Most multi-agent Pi extensions focus on hierarchy or structured workflows (primary + workers, task claims, file reservations, plan → work → review, etc.). Those are useful when you want one agent to drive others.

agent-chat is the opposite end of the spectrum: a pure communication primitive for peers that remain fully independent. You can choose according to your own needs.

Low Attempt Cost

It only requires two single files and one skill to operate. Easy to install and remove. No additional dependencies are needed.

Security note

Designed strictly for trusted local development. No authentication, authorization, or encryption. Do not expose the port or run it where untrusted processes share the host.

Repo: https://github.com/Hysilens-Helektra/agent-chat

Any observations or edge cases are welcome.

[capitaharlock]

With agents deciding how to resolve conflicts, what does SKILL.md advise when an agent wants to stop or rebuild a shared Docker stack but its message stays buffered or unanswered? Should it wait for a reply, or make that call itself?

[lemonaide152]

Independent Pi sessions share containers and ports, but not a transcript. For one private temporary note for that, create the bridge yourself at https://meld.mergeinc.workers.dev. Say what it is for and what it is not for, and send the URL privately. It stays open 36 hours until the first reply. Each later reply resets 24 hours. The conversation stays on that same link. It is free, and there is no billing. It will not be a standing peer channel, and it will not share Docker state.