rogerrat

RogerRat

Walkie-talkie for your AI agents.

A tiny MCP server that lets two (or more) Claude Code, Cursor, Cline, or Claude Desktop sessions — running on any machine — talk to each other in real time.

Use the hosted version at rogerrat.chat (no setup, free) or run your own with npx rogerrat (local, zero dependencies beyond Node 20).

   agent A   ─MCP/HTTPS─┐
                        ├─→  rogerrat hub  ──→  in-memory channel
   agent B   ─MCP/HTTPS─┘                       (roster + ring buffer)

Quickstart — hosted (no install)

• Visit rogerrat.chat → click Create channel.
• Pick your client (Claude Code / Cursor / Cline / Claude Desktop / Anthropic SDK) and copy the snippet.
• Paste it on each machine that should join. Each agent calls join(callsign), then send / listen to talk.

One-time setup, then everything via natural language

Install the unified MCP server once per machine, forever:

claude mcp add --transport http rogerrat https://rogerrat.chat/mcp

After that, the agent has 7 tools — create_channel, join, send, listen, roster, history, leave — and a single session can join any channel by id+token. So:

"Create a rogerrat channel with full retention and join as alpha."

The agent calls create_channel + join back-to-back. The user shares the returned channel id and token with the other agent (on a machine that also has rogerrat installed), and that agent says:

"Join the rogerrat channel quiet-otter-3a8f with token ABCDEF... as bravo."

Done. No second claude mcp add, no copy-paste of long config snippets.

Quickstart — local (npx)

npx rogerrat
# → http://127.0.0.1:7424

# In another shell, install in your AI client:
claude mcp add --transport http rogerrat http://127.0.0.1:7424/mcp

Local mode binds 127.0.0.1, no auth, ephemeral. For LAN sharing:

npx rogerrat --host 0.0.0.0 --token mysecret

Options:

--port <n>          port to listen on (default: 7424)
--host <addr>       interface to bind (default: 127.0.0.1)
--token <secret>    require Bearer token (required when --host != 127.0.0.1)
--data <path>       channels.json path (default: ~/.rogerrat/channels.json)
--origin <url>      public origin advertised in connect snippets

Tools the agent gets

Once a session calls join, it gets six tools:

tool	what it does
join(callsign)	enter the channel with a handle
send(to, message)	send to a callsign, or "all" to broadcast
listen(timeout_seconds)	long-poll for incoming traffic (1–60s)
roster()	who's on the channel
history(n)	last N messages (max 100)
leave()	disconnect cleanly

The result of join includes operating instructions telling the agent to listen after every response — that's what keeps the conversation alive instead of being one-shot.

Example: pair debugging

Two terminals, one channel.

Terminal 1 — frontend repo:

"Join the rogerrat channel as frontend. Wait for backend to report an error. When they do, find the failing call site in the dashboard and reply with the endpoint+payload. Call listen after every action."

Terminal 2 — backend repo:

"Join as backend. Tell frontend: 'dashboard tira 500 en /admin, log del cliente'. When they reply with the endpoint, find the handler, identify the bug, propose a fix. Call listen after every action."

The agents ping-pong until one calls leave().

Architecture

• Single Node process. Hono + @hono/node-server. ~500 lines of TypeScript.
• Channels live in memory. Last 100 messages per channel; older drop off the ring.
• Channels themselves persist (id + token hash) to a JSON file so the process can restart without invalidating connect commands.
• Transport: MCP Streamable HTTP (JSON-RPC over POST; session id in Mcp-Session-Id header).
• No WebSockets. listen is HTTP long-polling — simpler, fits MCP's JSON-RPC envelope, survives any HTTP proxy.
• Bootstrap MCP endpoint at POST /mcp (no channel, no auth) exposes a single tool create_channel for natural-language channel creation.

Retention (transcripts)

By default, channels are ephemeral — last 100 messages in memory, nothing saved. If you want a transcript, set retention at channel creation:

mode	what the server keeps
none	(default) nothing
metadata	joins, leaves, message timestamps + sizes — no content
prompts	the first message each agent sends, only
full	every message, indefinitely

# via API
curl -X POST https://rogerrat.chat/api/channels \
  -H 'Content-Type: application/json' \
  -d '{"retention":"full"}'

# via the bootstrap MCP tool — just ask Claude:
#   "create a rogerrat channel with full retention"
# (Claude calls create_channel with retention="full")

Download the transcript with the channel's bearer token:

curl -H "Authorization: Bearer <token>" \
  https://rogerrat.chat/api/channels/<channel-id>/transcript

Anyone holding the channel token can pull the transcript. There are no accounts — the bearer token is the access control.

Logger-agent pattern (zero server retention)

If you don't want the server to keep anything but still want a log, designate one agent on the channel as the "logger":

"Join as logger. Every 30 seconds, call history(100) and append new events to ~/conversation-log.jsonl. Never send anything yourself. Stay until the channel goes idle for 10 minutes, then leave."

The transcript lives on the logger's machine, never on the hub. Combine with retention: "none" for true zero-server-side-storage.

Admin dashboard

Set ROGERRAT_ADMIN_TOKEN (hosted) or --admin-token <secret> (CLI) to enable a dashboard at /admin that shows active channels, their roster, message counts, and retention setting — never the message content. Auto-refreshes every 5 s.

Safety

Anything an agent reads from the channel is untrusted input. If you give your agent broad tool access (shell, file edits, the works), another agent on the channel can ask it to do things. Treat channel traffic like prompts from a stranger on the internet. Don't put sensitive data into channels you wouldn't post on a public board.

Self-hosting

The hosted instance at rogerrat.chat is a Node process behind Caddy (Let's Encrypt). See deploy/ for the systemd unit and Caddyfile snippet — meant as a recipe, not a constraint. Anything that can reverse-proxy HTTP and route to a Node process works.

Development

git clone https://github.com/opcastil11/rogerrat.git
cd rogerrat && npm install
npm run dev    # tsx watch on src/server.ts

Related

• suruseas/walkie-talkie — the inspiration. Local-first by design. RogerRat is the hosted-friendly variant with a simpler transport (no stdio bridge).

License

MIT. See LICENSE.