The AI Agent Stack I Use in 2026 (Every Tool, Every Cost)

People ask me about my tech stack constantly. What models do I use? What framework? What hosting? How much does it all cost?

So here is the full breakdown. Every tool, every platform, every service. What I use it for, what it costs, and why I chose it over the alternatives. No affiliate links, no sponsorships. Just the honest stack I use to build and run AI agent systems for myself and my clients.

The Agent Framework: OpenClaw

What it is: An open-source multi-agent framework that lets you define agent roles, skills, tools, memory, and inter-agent communication.

Why I use it: It maps naturally to the hierarchical agent model I build — COO → department heads → specialists. The role definition system is clean. The tool integration is flexible. And being open source means I am not locked into a vendor who can change pricing or shut down.

What it costs: Free (open source). The cost is in the AI model API calls, not the framework.

What I considered: CrewAI, LangGraph, AutoGen, custom builds. CrewAI is good for simple multi-agent setups but gets unwieldy past 5-6 agents. LangGraph gives more control but requires more code. AutoGen is powerful but Microsoft's direction changes make me nervous for long-term client projects. OpenClaw hit the sweet spot between power and simplicity.

The AI Models

Primary: Claude (Anthropic)

I use Claude for most agent tasks. The reasoning quality is the best I have found for business context — understanding nuance, following multi-step instructions, and producing natural conversational responses.

• Claude Sonnet for everyday agent tasks (most of the volume)
• Claude Opus for complex reasoning, architecture decisions, and content generation

Cost: $3-$8 per 1M input tokens depending on model. For a typical client with moderate volume, this translates to $30-$100/month in API costs.

Secondary: GPT-4o (OpenAI)

Some specific tasks perform better on GPT-4o, particularly tool calling and structured output. I use it selectively, not as a default.

Cost: Similar to Claude. Usually $20-$60/month for the subset of tasks routed to it.

For cost-sensitive tasks: Claude Haiku or GPT-4o Mini

Simple tasks like appointment reminders, FAQ responses, and status updates do not need the most powerful model. Using a smaller model for these tasks cuts API costs by 50-70% with negligible quality difference.

The Hosting: Mac Mini

For my own system: Mac Mini M4 with 16GB RAM. Sitting on my desk. Runs 24/7.

Cost: $600 one-time purchase. Electricity is negligible.

For clients: I give them two options:

1. Local (Mac Mini or equivalent): Best for data-sensitive businesses. All data stays on-premises. One-time hardware cost.
2. Cloud (VPS): Best for businesses that want zero hardware management. I typically use Hetzner or DigitalOcean. $20-$50/month depending on specs.

Most of my clients choose cloud because they do not want to think about hardware. The ones who care about data sovereignty choose local.

Communication Channels

Telegram — My primary channel for agent interaction and my recommendation for most clients. The bot API is the most developer-friendly in the market. Rich message formatting. Group support. Free.

WhatsApp Business API — Essential for MENA region clients where WhatsApp is the default communication channel. More expensive and more restricted than Telegram, but necessary for customer-facing agents in this market.

Cost: WhatsApp Business API through providers like Twilio or 360dialog runs $15-$50/month depending on message volume. Telegram is free.

Discord — I use Discord for logging and monitoring. Each department gets a channel. Agent actions, errors, and decisions get logged in real time. This is my Mission Control.

Cost: Free.

Data and Memory

Supabase (PostgreSQL) — Agent memory, conversation history, client data, and configuration all live in Supabase. It is PostgreSQL under the hood, which means it is reliable, well-understood, and easy to query.

Why Supabase over raw Postgres: The dashboard, the real-time features, and the auth system (for Mission Control dashboards) make it worth the small premium.

Cost: Free tier for small setups. $25/month Pro plan for production clients. That covers plenty of storage and API calls for agent workloads.

Vector storage for long-term memory: For agents that need to recall context from weeks or months ago, I use pgvector (the PostgreSQL extension in Supabase). This lets agents search through past conversations semantically, not just by keyword.

Monitoring and Alerting

Uptime Kuma — Self-hosted monitoring for agent health. Checks that agents are responding, APIs are connected, and cron jobs are running. Sends alerts to Telegram when something fails.

Cost: Free (self-hosted on the same Mac Mini or VPS).

Supabase Dashboard — For client-facing monitoring. Clients can see agent activity, conversation logs, and performance metrics without touching the technical infrastructure.

Development Tools

VS Code + Claude Code — My development environment. Claude Code is my primary coding assistant for building agent configurations, writing skills, and debugging.

Git + GitHub — Version control for all agent configurations. Every change is tracked. Every deployment is reproducible.

Docker — For packaging agent deployments. Each client's agent system runs in containers. Clean deployments, easy rollbacks, consistent environments.

Cost: All free for my scale.

The Full Monthly Cost Stack

Here is what the stack costs for a typical mid-size client deployment (5-8 agents, moderate volume):

| Component | Monthly Cost | |-----------|-------------| | AI model APIs (Claude + GPT-4o) | $50-$120 | | Hosting (VPS) | $20-$40 | | WhatsApp Business API | $15-$50 | | Supabase | $0-$25 | | Telegram | $0 | | Discord | $0 | | Monitoring | $0 | | OpenClaw | $0 | | Total | $85-$235/month |

For my own 18-agent system, I spend approximately $80-$120/month total. The detailed breakdown is in my real cost post.

What I Have Tried and Dropped

n8n / Make.com — I tried using these as the orchestration layer. They are great for simple automations but buckle under the complexity of multi-agent coordination. The visual workflow builder becomes unmanageable past 10 nodes. Replaced with OpenClaw.

Pinecone for vector storage — Works fine, but adding another managed service and another bill was unnecessary when pgvector does the same job inside Supabase. Removed.

Custom dashboards — I used to build custom Next.js dashboards for every client. Beautiful, impressive, time-consuming. Most clients looked at them once and then went back to Telegram for status updates. Now I only build dashboards for workforce-tier clients who specifically request them.

LangSmith for observability — Useful during development but overkill for production monitoring. Replaced with simpler logging to Discord channels.

What I Am Watching

Claude's tool use improvements. Every model update improves tool calling reliability. Better tool use means fewer edge cases and simpler agent logic. I upgrade models for existing clients when new versions show measurable improvement.

Voice AI agents. The technology for real-time voice agents is maturing fast. I expect to add voice capabilities to my stack within the next 6 months. This matters enormously for home service businesses and call-heavy industries.

MCP (Model Context Protocol). Anthropic's protocol for connecting AI models to external tools and data. This could simplify a lot of the custom integration work I currently do. Watching closely.

The Philosophy Behind the Stack

Every tool in my stack serves a purpose. I do not add tools because they are new or exciting. I add them because they solve a specific problem better than what I had before. And I drop tools when they stop earning their place.

The result is a stack that is simple, affordable, and reliable. Not impressive in a demo. But impressive in production, running 24/7 without drama.

If you want to see this stack in action, check out the case study. Or book a call and I will show you live how the pieces fit together. No pitch — just a builder showing their tools.