People ask what tools I use. Fair question—if you’re going to take the “AI-assisted solo developer” path, you want to know what the actual setup looks like.
Here’s the honest answer: it’s simpler and cheaper than you’d think.
The Core Stack
Claude API / Anthropic
The AI brain. Claude Opus 4.5 for complex reasoning tasks. Claude Sonnet for faster, cheaper operations. The Claude Code CLI for the actual development interface.
This is the non-negotiable piece. The quality of the underlying model determines everything. Opus 4.5 is currently the best I’ve found for the kind of multi-step reasoning that software development requires.
Cost: Variable based on usage. Heavy development weeks run higher, light weeks are lower. The leverage makes it worth it.
Hosting: Mix of VPS and Managed Platforms
I use a combination:
- Hetzner for dedicated VPS hosting when I want full control and predictable costs. A cheap VPS handles surprising amounts of traffic.
- Railway for projects where I want managed infrastructure without thinking about servers.
- Vercel for frontend deployments where I want great DX and global CDN.
- Netlify for static sites and JAMstack projects.
The choice depends on the project. Client work that needs reliability gets Railway or Vercel. Personal projects and experiments often start on a cheap VPS.
Domain + DNS: Cloudflare
All domains route through Cloudflare. Free tier covers almost everything. DNS management, SSL, basic protection, reasonable analytics.
Database: Varies
- SQLite for single-server applications. Simpler than you’d think—handles more than people expect.
- PostgreSQL (via Railway or Supabase) when I need a “real” database.
- Plain JSON/JSONL files for simple logging and storage needs.
The Development Environment
Claude Code CLI
This is where I spend most of my time. The CLI interface to Claude, running in the terminal, with access to the filesystem, the ability to run commands, and the skills system that powers the 8-agent SDLC.
VS Code
Still use it for code review, navigation, and when I want a GUI. But increasingly, the primary “IDE” is the Claude Code terminal.
Git + GitHub
Version control is still version control. Nothing fancy here.
Docker
Local development and deployment consistency. docker-compose for multi-service setups.
What I Don’t Use
This list is as important as what I do use:
No Kubernetes. Docker-compose handles everything I need. When I need to scale, I either throw hardware at it or use a managed platform that handles orchestration.
No GitHub Actions for CI. Local test.sh script before deploy. GitHub Actions is great for teams; it’s overhead for solo work.
No Prometheus/Grafana. Cron job health checks + push notifications for alerting. Server metrics from the cloud provider dashboard when I need them.
No Terraform. I can set up a server by hand in 10 minutes. For most projects, that’s fine. When I need reproducibility, I write a setup script.
No Jira/Linear for project management. Markdown files. PROJECT-STATE.md tracks what’s in progress. Todo.txt for personal task management.
Why This Works
Complexity matches scale. I’m not a 50-person engineering org. I don’t need tooling built for 50-person engineering orgs.
AI handles the complexity. The whole “right-sized tooling” argument applies differently when you have AI assistance. Kubernetes is complex for humans but also complex for AI to work with. bash scripts are simple for humans AND simple for AI.
Managed services for the hard parts. I don’t want to manage databases or CDNs. That’s solved infrastructure. Pay for managed services and focus on the actual work.
Predictable costs. No surprise bills. I know roughly what I’m spending each month. VPS hosting is fixed cost. API usage is the variable, and I can monitor it.
What Might Be Missing
Honest assessment of limitations:
High availability. A single VPS isn’t redundant. If the server goes down, the service goes down. For critical applications, you’d want redundancy. For most of what I build, 99.5% uptime is fine.
Scale. This setup handles thousands of users, not millions. If something actually takes off, I’d need to adapt. That’s a good problem to have.
Team collaboration. This is a solo developer setup. If I had multiple engineers, we’d need CI/CD, better code review tooling, more structured environments. The simplicity comes from single-player mode.
Backup rigor. I should probably be more systematic about backups. Currently relying on VPS snapshots and git. Works, but isn’t enterprise-grade.
The Decision Framework
When choosing tools, I ask:
-
Does the complexity match my scale? If a tool is designed for problems I don’t have, it’s overhead.
-
Can AI work with it? Simple, obvious tools are easier for AI agents to operate reliably.
-
What’s the marginal cost of one more project? Infrastructure should feel cheap to expand, not expensive.
-
Can I debug it at 2am? If something breaks, can I figure out what’s wrong with my brain at 50% capacity?
-
What’s the exit cost? If this tool/platform goes away or becomes hostile, how painful is migration?
Most enterprise tools fail at least one of these for solo/small-team work.
For Those Starting Out
If you’re building your own AI-assisted development setup:
Start simpler than you think. A VPS with Docker and a Claude API key gets you surprisingly far.
Don’t pre-optimise. You can add complexity when you need it. You can’t easily remove complexity you’ve already built.
Track actual costs. Know what you’re spending. Unexpected bills suck and create stress that takes mental energy away from building.
Focus on the model. The AI capability matters more than the infrastructure around it. Invest in understanding and using the AI well before investing in elaborate tooling.
Closing Thought
The infrastructure that supports AI-assisted development is simpler and cheaper than you’d expect. The magic isn’t in the deployment pipeline or the monitoring stack—it’s in the AI capability and how well you direct it.
Keep the infrastructure boring. Keep the focus on building.
This is Part 8 of a series on AI-assisted software development. Previously: Requirements Are the New Code. Next: What Level 4 Actually Requires.