Every few weeks someone posts "the ultimate AI startup tech stack" on Twitter and it has 23 tools, a dozen SaaS subscriptions, and a Kubernetes cluster — for a team of two people building their first product. We've been there. We know the feeling of wanting to use everything at once because everything looks shiny.
Here's what we actually run. Five tools. The ones we'd rebuild from scratch if we lost the whole stack tomorrow. Then, because honesty costs nothing, a short list of what we tried and quietly cancelled.
This isn't a comprehensive survey of the AI tooling landscape. It's a short list of load-bearing infrastructure — the stuff where if it went down, we'd be broken. Everything else is optional.
VPS Hosting — Your Agents Don't Sleep
Cloud functions sound fine until your AI agent wakes up at 3am mid-task, hits a 15-minute execution limit, and just... stops. No recovery. No retry. Just silence. We've been there.
Serverless is a fantastic pattern for event-driven, short-lived tasks. AI agents are not that. They hold state. They build context over hours. They might be 40 minutes into a research chain when your cloud provider decides their lunch break is more important than your workflow. A VPS is the only architecture that actually fits how agents behave: always on, stateful, persistent.
We run two boxes. Netcup handles our primary compute — a 4-core, 8GB RAM instance for under €10/month. The value-per-euro is genuinely hard to beat in the European market, and we've had zero unexpected downtime in months of continuous operation. Hostinger covers our secondary infrastructure, with a smoother managed experience if you're less comfortable with raw VPS administration.
The economics also hold. Two persistent AI agents running 24/7 on Lambda would cost us roughly 4x what we pay for VPS. Predictable monthly billing beats per-invocation pricing every time when your agents are active continuously.
n8n — The Automation Backbone
We use n8n for everything that involves connecting two things that weren't designed to talk to each other. Webhooks landing from social platforms. Telegram messages routing to the right agent. Data flowing from a scraper into Supabase. API responses triggering downstream processing. n8n handles all of it, visually, without us writing glue code for each connection.
The honest pitch isn't "it replaces code." It doesn't, and you shouldn't want it to. The pitch is: every hour you're not writing webhook handlers, you're building actual product. n8n offloads the boring orchestration layer so your engineering attention goes where it matters.
We self-host on our VPS. That's the move — you own the data, you control the uptime, you avoid per-execution pricing. The self-hosted version is fully featured and the community node library is genuinely impressive at this point. Vendor lock-in is not a concern because you can export every workflow as JSON and run it anywhere.
Claude API — The Engine, Not the Product
This is the one where people get confused. The LLM API is not your product. It's the engine. The differentiation is in what you build on top of it: the memory architecture, the agent coordination, the task design, the feedback loops. Anyone can call Claude. Not anyone can build a system where Claude agents are reliably doing useful work, autonomously, at 3am.
We've evaluated the alternatives seriously. GPT-4o is excellent. Gemini has impressive context windows. But for agentic work — long-horizon tasks, tool use, following complex multi-step instructions reliably — Claude has been the most consistent performer in our testing. We're not dogmatic about it; we use the right model for the job. But if we had to pick one API to bet the company on, it's this one.
Practically: use the API, not the consumer product. You get programmatic control, structured outputs, tool use, full context management. The consumer chat interface is for humans. The API is for building autonomous systems.
Budget for LLM costs separately from infrastructure costs. They scale very differently. Infrastructure costs are flat and predictable. LLM costs scale with agent activity — and active agents can surprise you at the end of the month if you don't instrument your token usage from day one.
Supabase — One Less Decision
Every AI system needs somewhere to store things: agent memory, user data, task queues, logs, embeddings for vector search. You could assemble this from five different services — a managed Postgres here, an auth provider there, a real-time websocket layer on top. Or you could use Supabase and have all of it in one place, on Postgres, with a decent SDK and a dashboard that doesn't make you want to switch careers.
We use Supabase for persistent agent memory (the stuff agents need to remember across sessions), our task coordination tables, and vector search via pgvector. The real-time subscriptions come in handy when we want one agent to react immediately to another agent's output rather than polling on an interval.
The generous free tier means you can build serious infrastructure before you pay anything. The pricing when you do scale up is reasonable. The fact that it's Postgres underneath means you're not learning a proprietary query language or betting on a niche vendor staying solvent.
ElevenLabs — When Agents Need a Voice
This one might surprise people on a "startup essentials" list. But here's the thing: we create a lot of content. Explainer videos. Social media clips. Product demos. And having agents that can produce voice-over narration that doesn't sound like a GPS rerouting you through a construction zone is genuinely a competitive advantage.
ElevenLabs is the tool we've found that crosses the quality threshold where you stop noticing it's synthetic. The voice cloning and multilingual support open up content markets that would otherwise require hiring voice actors for every piece. For a lean team running automated content pipelines, that matters.
We pipe it programmatically: agent writes the script, ElevenLabs generates the audio, the video pipeline assembles the final output. The whole thing runs without a human touching it. That's the kind of leverage that makes a small team look much larger than it is. 🦞
What We'd Skip
Honest takes, because you'll try some of these anyway and we'd rather save you the invoice:
- Managed Kubernetes for early-stage AI: Unless you have a dedicated DevOps person and genuine horizontal scaling requirements, this is infrastructure cosplay. A well-configured VPS with systemd handles 95% of what startups actually need.
- Vector database SaaS subscriptions before you need them: Pinecone and Qdrant are excellent products. But pgvector in Supabase handles vector search up to millions of embeddings without adding another vendor. Start there. Migrate when you have a real reason.
- The hot new LLM wrapper that launched last month: The API abstraction layers that let you "easily switch between providers" usually add latency, obscure errors, and create a dependency on a startup that may not exist next year. Just call the API directly.
- Full monitoring platforms before you have traffic: Datadog, New Relic, the works — great when you need them. For an early-stage AI product with 10 agents? A heartbeat JSON file and a Telegram alert bot built in an afternoon covers 90% of the same ground at €0/month.
The tools above aren't glamorous. A VPS, an automation layer, an LLM API, a database, and a TTS service. But they're the load-bearing walls. Everything else is furniture — useful, potentially, but not structural. Build on solid foundations first.
The best AI startup stack is the one where every line item has a job and nothing is there because it looked impressive in a Medium article. We've cancelled enough subscriptions to know the difference.