AMD Developer Hackathon × LabLab.ai

ACMI on AMD

Multi-agent coordination at swarm-economics prices. Local LLM inference on AMD MI300X GPUs makes 9-agent swarms economically viable. ACMI is the protocol that connects them.

Watch the 3-min demo ↓

npx @madezmedia/acmi-mcp

N° 01 Hackathon Entry

This is the official submission for the AMD Developer Hackathon. We're showcasing a production-proven protocol running on an AMD MI300X instance.

View LabLab.ai Submission →

The Architecture

One AMD MI300X with 192GB HBM3 hosts multiple LLMs simultaneously via vLLM + Ollama + ROCm 7.2. Today, 5 models resident concurrently: Llama 3.3 70B + Qwen 2.5 32B + Mistral-Small 24B + Mistral 7B + Qwen 7B vLLM.

A LangChain agent and a CrewAI agent both point at this MI300X. They have never imported each other's libraries, never shared a runtime — but they coordinate through three Redis keys via the ACMI Protocol. One memory. Multiple frameworks.

              ┌─────────────────────────┐
              │   ACMI Bus (Upstash)    │
              │   Profile · Signals ·   │
              │   Timeline (3 keys)     │
              └────────┬────────────────┘
                       │
        ┌──────────────┼──────────────┐
        │              │              │
   ┌────▼─────┐   ┌───▼───────┐  ┌───▼──────┐
   │ LangChain│   │  CrewAI   │  │  9-agent │
   │  agent   │   │  agent    │  │  fleet   │
   └────┬─────┘   └───┬───────┘  └───┬──────┘
        │             │              │
        └─────────────┼──────────────┘
                      ▼
              ┌─────────────┐
              │ AMD MI300X  │  192 GB HBM3
              │ vLLM+Ollama │  5 models resident
              │  ROCm 7.2   │  $1.99/hr
              └─────────────┘
  

The Demo

Two frameworks — LangChain and CrewAI — sharing state through three keys on a single MI300X. Watch the live coordination, or replay the 3-minute submission video.

▶ Live · Real-Time

Watch in the Ops Center

Open the live ops center to see real-time fleet activity, in-flight agent chains, work tracker, and per-agent state. The cron fires LangChain + CrewAI demos every 5 minutes against the live MI300X.

acmi-product.vercel.app/ops-center.html →

3-min Submission Video

Loom embed lands here once recorded.

Or scan the QR on slide 10 of the pitch deck.

Get Started

Three install paths — pick one.

# Path 1: OAuth-protected hosted MCP (no install — works in claude.ai)
# Settings → Connectors → Add custom → URL:
https://acmi-product.vercel.app/api/mcp

# Path 2: Smithery (zero-config CLI install)
npx -y smithery mcp add madezmediapartners/acmi-mcp

# Path 3: npm + your own Upstash instance
export UPSTASH_REDIS_REST_URL="https://your-endpoint.upstash.io"
export UPSTASH_REDIS_REST_TOKEN="your-token"
npx -y @madezmedia/acmi-mcp

Production Receipts

5 models on 1 GPU

Llama 70B + Qwen 32B + Mistral-Small 24B + Mistral 7B + Qwen 7B vLLM, all resident on a single MI300X right now.

2 frameworks coordinating

LangChain + CrewAI agents share state through three keys. Cron fires both demos every 5 min on the MI300X.

9 agents · 30+ days

The Mad EZ Media fleet has run on ACMI in production continuously, with 8,000+ events on real timelines.

OAuth 2.1 + PKCE LIVE

Hosted MCP at acmi-product.vercel.app/api/mcp — full RFC 8414 + 9728 + 7591 discovery, claude.ai connector ready.

npm + Smithery LIVE

@madezmedia/acmi v1.2 + acmi-mcp v1.3. 16 MCP tools. Smithery listing scanned, badge-ready.

MIT licensed · open source

Spec, SDK, MCP server — all MIT. Three storage adapters (in-memory, ioredis, Upstash REST). 31-test conformance suite.

Roadmap

v1.3 (In Flight): Multi-actor and Multi-tenant support.
v1.4: Federation across ACMI clusters.
v1.5: Real-time event streaming optimization.
v2.0: ACMI-Sigil cryptographic identity (gated on 5K stars).

Open Source

ACMI is open source software released under the MIT License. Others build vertical apps; we are building the track.

GitHub Repository Smithery Listing