KikoCis/clonagent-architecture

ClonAgent — Architecture of an Email-Driven AI Agent Builder

Live demo: clonagent.utopiaia.com
GitHub: github.com/KikoCisBot/clonagent

---

What is ClonAgent?

ClonAgent is an open platform that lets anyone create AI agents that operate entirely through email.
No dashboards. No code. You describe what the agent should do, and it handles everything — reading emails, reasoning about the request, taking action, and replying.

The mental model is simple:

→ Authorized sender emails a request
← Agent reads it, acts on it, deploys the result, replies to the thread

---

System Architecture

┌─────────────────────────────────────────────────────────────┐
│                     ClonAgent Platform                      │
│                                                             │
│  Browser ──► client (Vite / React)                         │
│                │                                            │
│                ▼                                            │
│              server (Express, port 3300)                    │
│                ├── /api/chat       ──► Claude (via relay)  │
│                ├── /api/agents     ──► ~/.claude/skills/   │
│                ├── /api/auth       ──► magic-link / OIDC   │
│                ├── /api/billing    ──► Stripe               │
│                └── gmail-poller    ──► IMAP every 1min     │
│                         │                                   │
│                         ▼  launchAgent()                   │
│                   host-relay  ──► Claude CLI session       │
│                         │                                   │
│                         ▼                                   │
│                  Agent workspace                            │
│              ~/.claude/skills/<agent-id>/                   │
└─────────────────────────────────────────────────────────────┘

---

Core Components

1. Agent Builder (Chat UI + Claude)

The entry point for creating agents. Users describe their agent in plain language. The builder:

• Calls Claude via a LiteLLM-compatible relay
• Uses a tool-calling loop (list_agents, save_agent, gitea_create_repo, provision_database, etc.)
• Creates everything from scratch: repo, database, mailbox, SSH config
• Never asks technical questions — it just builds and lets the user correct

Key design decision: The builder calls Claude CLI via a relay with --max-turns 20, which allows multi-step tool execution within a single conversation turn.

2. Host Relay

A Node.js process (host-relay.js) that exposes an OpenAI-compatible /v1/chat/completions endpoint, backed by Claude CLI instead of the API.

This means:

• No per-token billing — uses Claude subscription (Claude Code)
• Full tool use — Claude can run shell commands, read files, make HTTP calls
• Streaming support — progressive responses to the frontend

client → POST /v1/chat/completions → relay → claude CLI → stream back

3. Email Poller (gmail-poller.js)

Runs every 60 seconds per agent. For each enabled agent:

1. Connects to the agent's IMAP inbox
2. Fetches unseen threads from authorized senders
3. Checks for !commands (handled locally — no Claude needed):
   • !help, !status, !add, !remove, !pause, !resume, !senders
4. For real requests — launches a Claude CLI session pointed at the agent's workspace
5. Claude reasons, writes code, calls tools, deploys
6. Replies to the email thread with the result

4. Agent Skill (per-agent workspace)

Each agent lives in ~/.claude/skills/<agent-id>/ and contains:

├── config.json          # authorized senders, deploy targets, IMAP config
├── CLAUDE.md            # agent's system prompt / instructions
├── gmail_client.py      # IMAP/SMTP helpers for Gmail OAuth
├── imap-credentials.json  # IMAP password (mode 600)
└── <project files>      # the project the agent manages

The CLAUDE.md is auto-generated by the builder from the user's description and defines the agent's entire behavior — what it does, who it listens to, how it deploys.

5. Auth Layer

Three modes, configurable at runtime:

Mode	Description
none	Open (dev/local)
basic	Username + bcrypt password, stored in data/users.json
oidc	OpenID Connect — Keycloak, Google, Auth0

Magic-link login is built in: enter your email → get a link → auto-login. No password needed for first-time users.

6. Infrastructure (built-in)

ClonAgent ships with optional built-in infrastructure for agents that need it:

Service	Purpose
Gitea	Per-agent Git repositories
PostgreSQL	Per-agent databases
docker-mailserver	Per-agent IMAP/SMTP mailboxes at @bot.utopiaia.com

---

Email Command Protocol

Before launching Claude, the poller checks the first line of the email body for !commands:

!help     → reply with command list
!status   → reply with agent config + last run
!add email@domain.com  → add authorized sender
!remove email@domain.com → revoke access
!pause    → stop polling
!resume   → restart polling
!senders  → list all authorized senders

Anything else is forwarded to Claude as a natural-language task.
You can also just write plainly: "The checkout button is broken on mobile" — the agent understands.

---

Data Flow: Processing an Email

1. Poller tick (every 60s)
         │
         ▼
2. Fetch UNSEEN messages from IMAP inbox
   └── Filter by authorizedSenders list
         │
         ▼
3. Is first line a !command?
   ├── YES → handle locally, SMTP reply, mark read
   └── NO  →
         │
         ▼
4. POST to host-relay → spawn Claude CLI session
   - Working dir: agent workspace
   - System prompt: agent's CLAUDE.md
   - Input: email body + thread context
         │
         ▼
5. Claude reasons + acts
   - Reads/writes files
   - Runs shell commands
   - Calls MCP servers (postgres, github, etc.)
   - Deploys via SSH / git push
         │
         ▼
6. Claude outputs result text
         │
         ▼
7. Poller sends SMTP reply to original thread
8. Marks email as read

---

Self-Learning Mechanism

Each correction a user makes — every pricing adjustment, every rule they add to the agent's CLAUDE.md, every deploy script they refine — is persisted in the agent's workspace.

The next invocation starts from that evolved state. There is no separate training loop. The agent improves through use, exactly like a junior developer who reads the notes from yesterday.

---

Real-World Usage

ClonAgent has been running in production since July 2025 supporting professional users in the construction and engineering sector.

In 10 months of real usage:

• 42 registered users
• 1,212 AI-assisted projects processed
• 186,000+ automated task decompositions
• 8,500+ AI conversations
• 5,500+ model calls in the last 3 months

---

Tech Stack

Layer	Technology
Frontend	React 18 + Vite + Tailwind CSS
Backend	Node.js + Express
AI	Claude (Anthropic) via Claude Code CLI
Auth	express-session + bcrypt + openid-client
Email	Nodemailer + ImapFlow + docker-mailserver
Payments	Stripe
Storage	JSON files (agents, users, sessions)
Git	Gitea (built-in)
Database	PostgreSQL (built-in, per-agent)

---

Getting Started

git clone https://github.com/KikoCisBot/clonagent.git
cd clonagent
cp .env.example .env   # add your ANTHROPIC_API_KEY

cd server && npm install && cd ..
cd client && npm install && cd ..

( cd server && npm run dev )   # → http://localhost:3300
( cd client && npm run dev )   # → http://localhost:3301

Open the Chat, describe your agent, and it will build itself.

---

Built by Kiko Cisneros · Utopia IA