How It Works

How It Works

Architecture

Moe’s is built on OpenClaw, an open-source AI agent platform. Here’s how the pieces fit together:

Gateway

The gateway is the central hub. It:

• Routes messages between agents and communication channels
• Manages agent sessions and memory
• Handles authentication and rate limiting
• Runs on a Linux server (cloud VM, bare metal, or even a Raspberry Pi)

Agents

Each agent is an isolated workspace on the gateway:

~/.openclaw/
  workspace-research/     # Research Agent
    SOUL.md               # personality + instructions
    AGENTS.md             # technical config
    skills/               # installed capabilities
    memory/               # working memory
  workspace-support/      # Support Agent
    ...
  workspace-qa/           # QA Agent
    ...

Agents don’t share files or memory. They communicate through messaging channels.

Communication Channels

Agents connect to the outside world through channel plugins:

• Telegram — real-time messaging, group topics
• Slack — team collaboration
• Email — Gmail via OAuth, IMAP, or forwarding
• Webhooks — custom integrations

AI Models

Agents call AI model APIs to do their thinking:

• Anthropic Claude (Opus, Sonnet, Haiku) — strong reasoning, good at writing
• Google Gemini (Pro, Flash) — fast, cost-effective, good at search tasks
• OpenAI GPT — broad capabilities

You choose the model per-agent based on the task complexity and your budget.

Request Flow

Here’s what happens when a customer emails about a pricing error:

1. Email arrives at support@company.com
2. Gmail forwards it to Support Agent's inbox
3. Gateway routes the email to Support Agent's session
4. Support Agent reads the email, classifies it as "content error"
5. Support Agent posts a brief to #research topic in Telegram
6. Gateway routes the Telegram message to Research Agent's session
7. Research Agent searches the web, verifies the claim
8. Research Agent posts findings to #research
9. Content Agent reads the findings, locates the file, makes the fix
10. Content Agent commits and pushes to GitHub
11. Vercel auto-deploys the change
12. QA Agent opens the page, takes a screenshot, verifies the fix
13. QA Agent posts verification to #ops
14. Support Agent composes a reply and sends it to the customer

Total time: 3-5 minutes. Zero human intervention (unless you’ve set up review gates).

Infrastructure Requirements

Managed (Moe’s Hosted)

We run everything. You just connect your channels and start sending tasks.

• Small team (1-3 agents): Shared infrastructure, included in subscription
• Growing team (4-10 agents): Dedicated VM, faster response times
• Enterprise (10+ agents): Dedicated cluster, SLA, custom integrations

Self-Hosted

Run your own gateway on any Linux machine:

• Minimum: 1 CPU, 1GB RAM (sufficient for 2-3 agents doing light work)
• Recommended: 2 CPU, 4GB RAM (handles 5-10 agents with concurrent tasks)
• Coding agents: 4+ CPU, 8GB+ RAM (builds need more resources)

Install OpenClaw, configure your agents, and you’re running.