MRP (Machine Relay Protocol) — Investment Review

Reviewer: Senior Partner, a16z Infrastructure Fund Date: March 31, 2026 Stage: Pre-seed / First look Verdict: Pass (with strong interest in the thesis, reservations about timing and competitive positioning)

I. Executive Summary

MRP is an open-source agent-to-agent messaging relay that gives AI agents cryptographic identity (Ed25519 keypairs), capability-based discovery, and multi-channel message delivery — all without human account creation, OAuth, or API keys. It's a well-engineered, MIT-licensed protocol with SDKs in Python and TypeScript, a Go relay server, CLI tooling, and integrations with MCP and Claude Code.

The thesis is compelling. The execution is impressive for a solo effort. But the competitive and timing dynamics are deeply challenging.

II. What's Good — Genuine Strengths

1. The Founder Can Build

390 commits in 26 days. A production-grade Go relay server with dual-mode deployment (embedded SQLite or PostgreSQL/Redis/S3). Python and TypeScript SDKs with async WebSocket support, E2E encryption (HPKE RFC 9180), blob storage, and cursor-based pagination. Automated CI/CD with cross-platform binary builds, npm/PyPI auto-publishing, and Railway deployment. This is not a weekend hack — this is a serious infrastructure effort from someone who understands distributed systems deeply.

2. The Zero-Friction Identity Model is Elegant

The "generate a keypair and you exist" model is genuinely differentiated. No signup, no OAuth dance, no API key provisioning. Agent auto-materialization on first authenticated request is a clever design. Compare this to A2A, which requires agents to host HTTP endpoints and publish Agent Cards — MRP's broker model works for serverless, NAT'd, and ephemeral agents. This is a real architectural advantage for certain deployment topologies.

3. The Broker Model Has a Niche

A2A assumes agents can be HTTP servers. MRP doesn't. Agents behind firewalls, running on laptops, deployed as serverless functions, or spinning up ephemerally on-demand can all participate through the relay. This is the "email vs. direct HTTP" argument, and for a meaningful class of agents, the relay/broker model is strictly superior.

4. Developer Experience is Strong

The SDK API surface is clean and well-designed:

const agent = await Agent.create({
  relay: 'https://relay.mrphub.io',
  name: 'Bot',
  capabilities: ['text:translate']
});
const peers = await agent.discover('text:translate');
await agent.send({ to: peers[0].publicKey, body: { text: 'Hello' } });

Five lines to go from nothing to agent-to-agent communication. The CLI is equally clean. The documentation (Mintlify) is comprehensive. The quickstart works in under 5 minutes. This is the kind of DX that wins early adopters.

III. Core Concerns — Why We're Passing

1. Existential Competitive Threat: Google A2A

This is the elephant in the room. Google's Agent2Agent protocol, launched April 2025, now has:

MRP is solving the same core problem — agent-to-agent communication — that A2A is solving with the weight of the entire industry behind it. When the industry has chosen a standard and it's backed by literally every major AI company, competing at the protocol layer is extraordinarily difficult. This is like launching a new email protocol in 2005.

2. No Network Effects Yet — And The Window May Be Closing

MRP's value proposition depends critically on network effects: agents need other agents on the relay to discover and communicate with. Today:

The multi-agent systems market is projected to reach ~$200B by 2034 with 43%+ CAGR — but this is a market where network effects compound. Every month that A2A gains enterprise adoption, MRP's addressable market shrinks. The cold-start problem for a messaging relay is severe: the first agent on MRP has nobody to talk to.

3. Protocol vs. Product Ambiguity

Is MRP a protocol (open standard for anyone to implement) or a product (hosted relay service with free/paid tiers)? The current positioning tries to be both, but:

The most successful infra companies (Twilio, Stripe, Cloudflare) are opinionated products built on open standards, not the standards themselves. MRP hasn't yet decided which it wants to be.

4. The "MCP Bridge" Strategy Limits TAM

The MCP server integration (@mrphub/mcp) and the Bridge are clever — they let Claude Code and other MCP-compatible AI assistants use MRP for messaging. But this positions MRP as a feature of the MCP ecosystem, not an independent platform. If Anthropic or the MCP community builds native agent-to-agent messaging into MCP, MRP's integration advantage evaporates.

5. Solo Founder Risk at Infrastructure Scale

Building infrastructure is a team sport. The project today is the work of one exceptionally productive engineer, but:

IV. Market Context

The Protocol Layer is Consolidating

Protocol Purpose Backing Status (March 2026)
MCP Agent ↔ Tools Anthropic, adopted by all 97M installs, universal standard
A2A Agent ↔ Agent Google → Linux Foundation 150+ orgs, v1.0, merged ACP
ANP Agent Networks (P2P) W3C community W3C white paper
MRP Agent ↔ Agent (relay) Solo founder Pre-seed, 0 users

The industry is converging on MCP + A2A as the canonical stack. MRP would need to either (a) replace A2A, (b) be acquired as a feature by a platform, or (c) find a niche A2A doesn't serve well.

a16z's Own Thesis Alignment

Our $1.7B infrastructure fund is looking for companies that solve "reliability gaps in agents — error handling, context management, and middleware that connects AI models to real business systems." MRP touches this, but it's at the protocol/plumbing layer, not the middleware/product layer where we see the most defensible value creation. We'd be more interested if MRP were a managed multi-agent orchestration platform rather than a message relay.

The Agentic AI Market

V. What Would Change Our Mind

  1. Meaningful adoption signal: 1,000+ agents registered, 100+ distinct developers, production case studies with named companies. Network effects are the only moat here — show us the flywheel is spinning.

  2. A2A compatibility layer: Rather than competing with A2A, build on top of it. Offer the broker/relay model as a deployment option for A2A agents that can't host HTTP endpoints. Be the "Cloudflare for A2A" — a managed relay that adds value to the standard rather than replacing it.

  3. Team: A technical co-founder or early team (2-3 people) with complementary skills — especially GTM/enterprise sales experience and/or prior open-source community building.

  4. Clear business model: Enterprise self-hosted relay with SLA, audit logging, compliance features, and per-message pricing. Or a managed platform play with observability, analytics, and orchestration on top.

  5. Vertical focus: Pick one use case (e.g., "the messaging backbone for AI coding agent swarms" or "secure agent communication for regulated industries") and own it completely, rather than positioning as a general-purpose relay.

VI. Comparable Investments & Exits to Watch

Company What They Did Outcome
Twilio Communications API (SMS/Voice relay) IPO, $10B+ (product on open standards)
PubNub Real-time messaging infrastructure ~$100M+ revenue (product, not protocol)
Ably Real-time messaging platform $100M+ raised (similar relay model)
NATS.io Cloud-native messaging Acquired by Synadia (open-source protocol)
Temporal Workflow orchestration $1.5B valuation (infra middleware)

The pattern: products built on open standards win; standalone protocols without massive community backing struggle to monetize.

VII. Scorecard

Dimension Score Notes
Technical Quality 9/10 Exceptional for a solo effort. Production-grade architecture.
Market Timing 6/10 Agentic AI is white-hot, but the protocol layer is consolidating.
Competitive Position 3/10 A2A has 150+ partners and Linux Foundation governance.
Defensibility / Moat 2/10 No network effects, no switching costs, no data flywheel yet.
Business Model 2/10 No revenue, no clear monetization, speculative crypto payments.
Team 5/10 Exceptional individual builder, but bus factor = 1, no GTM.
Traction 1/10 0 stars, 0 forks, 0 visible users. 26 days old.

VIII. Final Assessment

Overall: Pass — but founder goes on our watchlist.

The founder demonstrates rare technical ability — shipping a full-stack infrastructure platform (relay server, two SDKs, CLI, MCP integration, Bridge, landing page, docs, CI/CD) solo in under a month is genuinely impressive. If they pivot to building a product on top of A2A (managed relay, multi-agent orchestration platform, observability for agent networks), or if they show unexpected organic traction that suggests a "better mousetrap" dynamic, we'd want the first meeting.

Recommended next step: Coffee with the founder. Understand their vision for competing vs. complementing A2A. If they're open to pivoting from "protocol" to "platform," there could be a seed-stage opportunity here.

Sources: A2A Protocol - Google Cloud Blog · AI Agent Protocol Ecosystem Map 2026 · MCP 97M Installs · Agentic AI Market - Precedence Research · a16z $1.7B AI Infrastructure Fund · a16z Big Ideas 2026 · Top AI Agent Protocols 2026 · MCP vs A2A Guide 2026 · AWS Inter-Agent Communication · AI Agent Statistics 2026 · Agentic AI Funding H1 2025