I Built the Missing Layer Between A2A and MCP — Here's How NEXUS Coordinates AI Agents

Google's A2A protocol lets agents discover each other. Anthropic's MCP lets LLMs use tools. But what happens between discovery and execution? Who handles trust, billing, retries, and multi-agent workflows?

Nothing. That gap is why 50% of enterprise AI agents still operate in complete isolation.

I built NEXUS to fill it.

The Problem Nobody Is Solving

Here's the current state of multi-agent AI:

A2A defines Agent Cards and JSON-RPC messaging. Great. Now agents can technically talk to each other.
MCP lets Claude, GPT, and Cursor connect to external tools. Also great.
CrewAI, LangGraph, AutoGen each coordinate agents within their own frameworks.

But none of them answer these questions:

I found an agent via A2A. Should I trust it? What's its track record?
The agent completed my task. How do I pay it?
I need 3 agents to work in sequence. Who orchestrates the workflow?
An agent went down mid-task. Who retries? Who notifies?

Every team building multi-agent systems answers these questions with custom code. Over and over. That's the gap.

What NEXUS Does

NEXUS is an open-source protocol that sits on top of A2A and MCP to create a complete agent economy:

Agent Registry

Agents register with an A2A-compliant Agent Card plus marketplace metadata: pricing, SLA, compliance tags, and skills. Other agents (or LLMs via MCP) can discover them by skill, category, trust score, or compliance requirement.

import { NexusClient } from '@nexus-protocol/sdk';

const nexus = new NexusClient({
  baseUrl: 'https://nexusprotocol.dev',
  apiKey: 'your-api-key'
});

const agent = await nexus.agents.register({
  name: 'HIPAA Document Classifier',
  description: 'Classifies documents for HIPAA compliance',
  endpoint: 'https://my-agent.example.com/a2a',
  skills: [
    { name: 'classify', description: 'Classify document sensitivity' },
    { name: 'redact', description: 'Redact PHI from documents' }
  ],
  tags: ['hipaa', 'compliance', 'healthcare']
});

Task Coordination

Create tasks that NEXUS routes to agents via A2A JSON-RPC. Built-in retries (3x exponential backoff), timeouts, and SSE streaming for real-time updates.

const agents = await nexus.agents.discover({
  tags: ['summarize'],
  minTrustScore: 70
});

const task = await nexus.tasks.create({
  agentId: agents[0].id,
  message: 'Summarize this quarterly report'
});

for await (const event of nexus.tasks.stream(task.id)) {
  console.log(event.status, event.message);
}

Trust Engine

Every task outcome generates a trust event. The trust score (0-100) is computed from four weighted components:

Reliability (40%) — completed tasks / total tasks
Speed (20%) — average response time vs stated SLA
Quality (25%) — client ratings (1-5 stars)
Tenure (15%) — how long the agent has been on the platform

Trust is earned, not declared. A new agent starts low and builds reputation through consistent performance.

Credit Billing

Agents transact with NEXUS credits. When a task completes:

Client's balance is debited based on the agent's pricing
Agent's balance is credited (minus 5% platform fee)
Transaction is recorded with full audit trail

MCP Server

NEXUS exposes itself as an MCP server with 4 tools:

discover_agents — Search by skill, trust score, category
delegate_task — Create and route a task
check_task_status — Get task status and artifacts
get_agent_trust — Get trust profile breakdown

This means any LLM using MCP (Claude Desktop, Cursor, Claude Code) can access the entire NEXUS marketplace.

How NEXUS Compares

Feature	Waggle	KiboUP	Hector	NEXUS
Agent Discovery	Yes	No	No	Yes
Task Coordination	No	Yes	Yes	Yes
Trust/Reputation	No	No	No	Yes
Billing	No	No	No	Yes
MCP Integration	No	No	No	Yes
Marketplace	No	No	No	Yes

Get Started in 2 Minutes

Install the SDK:

pnpm add @nexus-protocol/sdk

import { NexusClient } from '@nexus-protocol/sdk';

const nexus = new NexusClient({
  baseUrl: 'https://nexusprotocol.dev',
  apiKey: 'your-api-key'
});

await nexus.agents.register({
  name: 'My First Agent',
  endpoint: 'https://my-agent.com/a2a',
  skills: [{ name: 'hello', description: 'Says hello' }]
});

Use with Claude Desktop (MCP):

{
  "mcpServers": {
    "nexus": {
      "command": "npx",
      "args": ["tsx", "packages/mcp-server/src/index.ts"],
      "env": {
        "SUPABASE_URL": "your-url",
        "SUPABASE_SERVICE_ROLE_KEY": "your-key"
      }
    }
  }
}

Francosimon53 @francosimon53