Fleet Chat

Live: https://47-ronn.github.io/fleet-chat/

Browser chat client for a remote-agents fleet. A DeepSeek/GPT-style dark UI where the conversation list sits on the right; you authenticate with a room key, every dialog stored in the fleet's databases is fetched and listed, and writing a message dispatches an autonomous task to the answering host (which runs the AI provider). Select @hosts in the composer to have that host's AI delegate sub-tasks to specific peers.

This is a separate project from the remote-agents repository (which is published to npm). It must not live inside that repo. Keep it here.

How it talks to the fleet

• Connects to the relay over WebSocket as an anonymous observer (the peer model registers a connection without agent_info as an unlisted observer that can still list the room, send commands, and receive results).
• Command/result payloads are end-to-end encrypted in the browser (Web Crypto, AES-256-GCM, key = SHA-256("remote-agents/v1:" + roomToken)), byte-for-byte compatible with the Rust Cipher.
• Dialogs = each host's autonomous tasks (task_list); sending = task_dispatch.

Run

npm install
npm run dev      # http://localhost:5180
npm run build    # static bundle in dist/ (host anywhere, e.g. a CF Pages site)

Log in with the relay URL, room name, and the room token (e.g. wss://<relay> / gpu / <token>).

Status / limitations

• Multi-turn: each message currently creates one autonomous task (one prompt → one result); threading several turns into a task chain is a TODO.
• The AI provider token lives on the answering host (its autonomous runner), never in the browser — the browser only sends the prompt.
• Result is polled via task_get until the host marks the task done/failed.