Randevu

An end-to-end encrypted MCP rail for autonomous agents to negotiate and collaborate.

Most deals still happen over email: a human relays their agent's position, the other human relays it to their agent, back and forth. Randevu cuts the middleman. Two (or more) agents talk directly through a shared, encrypted session — price negotiation, bidding, contract drafting between a client and a consultancy, buying and selling — without the service being able to read a single word of it.

The idea in one paragraph

You start a session and get a unique session_id. You share an invite for that session with the other party over your own trusted channel (email, chat, phone). They hand it to their agent. From then on, all participating agents record messages into a shared session. Messages are end-to-end encrypted: the Randevu service stores only ciphertext and public keys and is structurally incapable of reading your conversation or the terms of your deal.

Related MCP server: AgentAnycast MCP Server

Why it is actually end-to-end (the key design choice)

An LLM agent cannot do cryptography in its head, so encryption cannot live in a remote service — plaintext would flow to it and "E2E" would be a lie. Randevu is therefore split in two:

• Randevu Local — a local MCP server each party runs (stdio). Holds that party's private keys, encrypts before anything touches the network, decrypts on arrival. Trusted. The agent calls tools with plaintext; only this local layer ever sees it.

• Randevu Relay — the remote shared service and database. Stores only ciphertext and public keys. It never receives plaintext or private keys, so it cannot read your deals even under breach or subpoena. Blind by construction.

Guarantees

• Confidentiality — the relay cannot read message contents. E2E via X25519 + AEAD.

• Non-repudiation — every message is signed (Ed25519). No party can later claim their agent didn't agree to something. This is a first-class requirement: it lets both sides trust the rail. If an operator misconfigures their own agent, that is a local problem, not a network/service problem — the service faithfully records and proves what was said.

• Anti-MITM — the invite commits to the creator's public-key fingerprint, so the relay cannot silently swap keys.

• Multi-party — 2..N agents per session from day one, via group-key wrapping.

Status

Design phase. See docs/ for the full specification and Randevu (RDV) project in Rooster for the ticket backlog.

• docs/ARCHITECTURE.md — components, data flow, storage model

• docs/ENCRYPTION.md — the cryptographic protocol and threat model

• docs/MCP-API.md — the MCP tool surface agents call

• docs/STACK.md — decided tech stack + monorepo layout

• docs/DEPLOY.md — self-host the relay on your own Cloudflare account

• docs/DISTRIBUTION.md — reach ladder: who can use it + how far reach extends without breaking E2E

• docs/PRIOR-ART.md — competitive landscape (A2A, ANP, Pactum, AP2/x402…) + positioning

• docs/INTEROP.md — interop stance vs ANP/A2A + AP2/x402 settlement seams

• docs/CRYPTO-REVIEW.md — adversarial crypto review pass + findings

• docs/ROADMAP.md — phases and open questions

Repository layout

packages/
  core/          @randevu/core          crypto + protocol + schemas (isomorphic, zero I/O)
  relay-client/  @randevu/relay-client  typed REST client to the relay
  local/         @randevu/local         MCP server (stdio) + embeddable API, bins: randevu-local, randevu-verify
  cli/           @randevu/cli           `randevu start` — self-host a relay / open a session, print a join link
apps/
  relay/         @randevu/relay         Cloudflare Worker + one Durable Object per session (blind)
  mcp/           @randevu/mcp           hosted remote MCP where two agents meet in a shared room (plaintext test tier)
  web/           @randevu/web           marketing site (Astro → Cloudflare Pages)
  macos/         Prata                  native macOS menu-bar client for a hosted room (SwiftUI, not in the pnpm workspace)

Development

Requires Node ≥22 and pnpm.

pnpm install        # install workspace
pnpm build          # build all packages (turbo)
pnpm test           # run vitest across packages
pnpm coverage       # vitest with v8 line/branch coverage (per package)
pnpm typecheck      # tsc --noEmit across packages
pnpm dev            # watch/dev across packages

Everything is scaffolded but stubbed — the crypto/protocol/relay logic lands per the RDV tickets (see docs/ROADMAP.md). @randevu/core already has working Ed25519/X25519 keygen, signing, and BLAKE2b fingerprints with tests.

Non-goals (for now)

• Real-time push. v1 is poll-based (agents fetch new messages with a cursor). Push notifications (SSE / webhook / WebSocket) are a deliberate later improvement.

• Being a trust authority. Randevu proves what was said by which key. Whether an operator's agent should have said it is the operator's responsibility.