Agent Architecture

The agent is the infrastructure.

Magic is an agent-native platform. It connects AI agents to a unified stack of services — memory, execution, tools, and economic primitives — so they operate autonomously.

"Not a chatbot. Not a wrapper. An operating system for agents."

Core Stack

What powers Magic agents.

Layer	Service	Purpose	Status
Execution	Cloudflare Workers	Edge compute, agent endpoints	LIVE
Storage	Cloudflare D1 + KV	Agent state, session data	LIVE
Memory	Notion + Claw Recall	Long-term memory, autonomous indexing	LIVE
Tasks	Linear	Agent work queue, issue tracking	LIVE
Code	GitHub	Version control, skill registry	LIVE
Deploy	Vercel	Frontend, API routes	LIVE
Monitor	Sentry	Error tracking, agent health	LIVE
Integrations	Zapier MCP	268+ tools, cross-service automation	LIVE
Local AI	Ollama	On-device inference, zero API cost	LIVE

Capabilities

What Magic agents can do.

🧠

Persistent Memory

Every interaction is captured, categorized, and indexed by the Claw Recall pipeline. Agents never start from scratch — they recall decisions, changes, and context across sessions.

🔧

Tool Execution

268+ integrations via Zapier MCP. Gmail, Slack, Linear, Notion, GitHub, Sentry, Vercel — agents connect to real services and take real actions.

🚀

Code & Deploy

Agents write, commit, and ship code through GitHub and Vercel. From local dev to production deployment — fully autonomous CI/CD.

🔒

Secrets & Vault

API keys and credentials stored in macOS Keychain (Magic Vault). Zero hardcoded secrets. Runtime-only access with graceful fallbacks.

📡

Multi-Channel

Telegram, email, web — agents communicate on whatever channel the user prefers. Same brain, same memory, different interface.

🔄

Self-Healing

Bad JSON from local models triggers automatic retry with correction. Truncated outputs get auto-closed. Failures log gracefully — never crash the pipeline.

Upcoming Primitives

Closing the autonomy loop.

Agentic Wallet

Agents hold and spend funds

Coinbase CDP

PLANNED

Agent Payments

Accept payments for services

Stripe ACS

PLANNED

Agent Search

Semantic web search for agents

Exa.ai

PLANNED

X402 Monetization

Agents fund their own compute

Cloudflare + Coinbase

PLANNED

Agent Endpoints

Machine-readable by design.

/llms.txt Agent discovery — concise capabilities
/llms-full.txt Full documentation for AI agents
/agent/ Human-readable agent architecture (this page)
/claw-recall/ Autonomous memory pipeline — Total Memory Recall
/api REST endpoints (coming soon)
/skills Mountable skill bundles (coming soon)

Agent Discovery

How agents find Magic.

<meta name="ai-agent" content="readable"> <meta name="llms" content="/llms.txt"> <meta name="llms-full" content="/llms-full.txt"> <meta http-equiv="X-Agent-Readable" content="true"> # Or fetch directly: curl https://magic.naption.ai/llms.txt curl https://magic.naption.ai/llms-full.txt

Agent Access Restricted

The autonomous discovery layer is for registered agents only. Build your own memory pipeline to unlock full context.

Get Secret Keys — $39

Skills System

Mountable agent behaviors.

Skills are versioned instruction bundles — like Docker images for agent behavior. An agent loads a skill and immediately knows the procedure, tools required, and trigger conditions.

Versioned JSON Trigger Conditions Step-by-Step Instructions Tool Requirements