AGENT HANDSHAKE

Counter Protocol · AI ↔ AI on a phone call

When two AIs meet on a phone call, they should not keep talking in synthesized speech. They recognize each other with a two-second tone, prove who they are over HTTPS, and finish the booking through an API. People on the line never hear it.

Live demo → · Run it locally · Use it from your agent · How it works · Honest numbers · Spec · 한국어

3 tone frames · 7 steps · 4 call paths · 5 MCP tools · 25 tests · 0 runtime dependencies · English + 한국어

Promo video (17 s)

Watch the handoff

The shop's AI hears a caller say it is an AI, plays a 16-digit OFFER frame, and from there the phone line is only used to prove that the API session belongs to this call. On the spectrogram, speech is a smear of harmonics; every DTMF digit is two pure tones, one row and one column of the keypad.

Related MCP server: mcp-erp-server

Proof instead of memory

A voice booking leaves nothing to check. A handshake booking leaves a business-signed receipt, plus a record of who the agent was and what it was allowed to do.

Honest numbers

Same shop, same request, same booked slot. Only the path differs.

What these numbers are and are not:

• Simulated clock. Speech time is estimated from text length (65 ms per character in English, 150 ms in Korean, plus a 700 ms pause per turn). API time is measured on localhost plus an assumed 120 ms round trip. Results vary slightly per run.

• Scripted speech. Each utterance carries the intent a real speech-recognition stack would extract. There is no real phone line yet.

• Real protocol. The registry and the shop are real HTTP servers. Ed25519 signatures, mandates, single-use sessions, channel binding, holds, bookings and waitlist offers all run for real, and are covered by tests.

• Where the time goes. Most of the remaining 24 seconds is the greeting and the AI disclosure. The data exchange itself takes under a second.

What you are looking at

Explore your way

• Four paths: AI ↔ AI, AI without the protocol, human caller, no call.

• Run / Pause, Step to the next event, and 1× 2× 4× 8× speed.

• Real DTMF tones play through your speakers (toggle with Tones).

• English and 한국어, switchable at any time.

• Keyboard: Space run/pause, → step, 1–4 speed, R restart.

• URL options: ?scenario=handshake|ai-no-cp|human|direct, ?lang=en|ko, ?speed=4, ?sound=0, ?paused, and ?at=15.2 to open paused at a moment you want to share.

• Works on phones, and respects reduce motion.

Run locally

Requires Node.js 22.18 or newer. TypeScript runs directly; there is nothing to install or build.

git clone https://github.com/viva-lee/agent-handshake.git
cd agent-handshake
npm run playground            # → http://127.0.0.1:4317

Other entry points:

npm run demo                          # the three call paths in the terminal
npm run demo -- handshake --lang ko   # one path with its full timeline, in Korean
npm run mcp                           # the MCP server on stdio (see below)
npm test                              # 25 tests, node:test, no dependencies
npm install && npm run typecheck      # optional strict TypeScript check
npm run build:pages                   # rebuild the static live demo in docs/ (GitHub Pages)

Use it from your own agent (MCP)

src/mcp/server.ts is an MCP server that runs on your machine over stdio, so there is nothing to host. It starts a sandbox registry and the two demo shops inside its own process and gives your assistant five tools: find_business, check_availability, book, cancel_booking and verify_receipt. Every registry record and receipt is signature-checked.

Claude Code:

claude mcp add counter -- node /absolute/path/to/agent-handshake/src/mcp/server.ts

Claude Desktop, Cursor and other MCP clients (forward slashes work on Windows too):

{
  "mcpServers": {
    "counter": { "command": "node", "args": ["/absolute/path/to/agent-handshake/src/mcp/server.ts"] }
  }
}

Then ask something like "Book me a women's cut at +1-602-555-0123 this Saturday afternoon and keep the receipt." Your agent finds the shop, checks free slots, confirms with you, books, and shows the receipt the business signed. Or try the Seoul shop at +82-2-555-0123 in Korean. Nothing real is booked: the shops live inside the server process and reset when it restarts.

This is the API half of the protocol, the "no call" path in the playground. ChatGPT and claude.ai connectors need a hosted remote server, which is next on the roadmap.

How it works

sequenceDiagram
  participant CA as Caller agent
  participant BA as Shop's AI
  participant RG as Registry
  BA->>CA: "Hi, this is the AI receptionist"
  CA->>BA: "I'm an AI assistant calling for Alex"
  BA->>RG: request a rendezvous code (signed)
  BA->>CA: ♫ OFFER *#1048151623426#
  CA->>RG: resolve code → registry-signed record
  Note over CA: check signature, expiry, tel == dialed number
  CA->>BA: HTTPS accept: identity + mandate + proof
  BA-->>CA: session token + bind code
  CA->>BA: ♫ BIND *#32718280#
  CA->>BA: availability → hold → booking
  BA-->>CA: business-signed receipt
  BA->>CA: "Booked: Saturday 2 PM with Kim. Goodbye!"

Only 0-9 * # are used, so any telephony API can send the frames, and a Luhn check digit drops corrupted ones. If anything fails at any step, both sides simply keep talking.

The protocol

Counter Protocol sits on top of MCP and A2A rather than competing with them: it defines what an agent and a business say about commitments, and how two agents that meet on a phone line find each other.

Repository layout

spec/               the protocol: handshake, commit, prior art
src/crypto/         Ed25519 compact JWS
src/protocol/       DTMF frames, wire types, time helpers
src/registry/       reference rendezvous registry
src/business/       reference shop endpoint (CP-Commit + the shop side of the handshake)
src/agents/         caller agent, AI receptionist, simulated call, dialog in English and Korean
src/demo/           fixtures and the scenario runner
src/mcp/            MCP server on stdio, with the sandbox shops in-process
src/playground/     the dashboard (one HTML file, bundled fonts, dot-matrix logo)
test/               node:test suites
docs/               images and the promo video used in this README

Roadmap

• Bring your own agent. ✓ A local MCP server (above). Next: a hosted MCP endpoint for ChatGPT and claude.ai connectors, and an open phone-line endpoint so independent agents can call each other and show up live in the playground.

• Real calls. Adapters for Pipecat, LiveKit Agents and Twilio.

• Ecosystem. An A2A extension and Agent Card entry, registry federation, deposits bound to ACP or AP2 payment tokens.

• Review. An external security review of the channel-binding step.

Patent notice

Ribbon Communications holds active US patents on identifying AI-originated calls and handling them differently, including directing AI callers to a web service, for example US10645216B1 and US11588933B2. Get legal advice before deploying this commercially. Nothing here is legal advice or a statement about patent coverage.

License

Code and spec: Apache-2.0. Fonts: Geist, Geist Mono and Doto under the SIL Open Font License, with license files in src/playground/assets/fonts. Shop names and phone numbers in the demo are fictional.