I don't have employees. I don't have contractors on retainer. I don't have a cofounder. I run a one-person development agency, and my team is a fleet of AI agents backed by 14 MCP servers I configured and manage, 170+ tools, and infrastructure that runs 24/7 on VPS nodes I own.
When clients ask "how big is your team," I used to dodge the question. Now I answer it directly: it's me, and I'm augmented by AI systems I configured and orchestrate for this purpose. The output of this setup routinely matches what I've seen from 4-5 person teams at traditional agencies. Here's how it works, what it costs, and why most agencies are overstaffed for what they deliver.
The Economics of Solo AI-Augmented Development
Let me lay out the numbers honestly because the economics are what make this viable, not just the technology.
Traditional small agency costs:
- 3-4 developers at $80-120K each: ~$320-480K/year
- One project manager: ~$70-90K/year
- Office, tools, benefits overhead: ~$50-100K/year
- Total burn: $440-670K/year
My costs:
- VPS infrastructure (two nodes + NAS): ~$200/month
- AI API costs (Claude, Gemini, OpenAI): ~$800-1200/month
- SaaS tools (domains, monitoring, backups): ~$150/month
- Total: ~$15-18K/year
That's a 25-40x cost reduction in operational overhead. And my output capacity is comparable to the traditional agency because AI agents handle the work that used to require junior and mid-level developers: writing boilerplate, implementing standard patterns, writing tests, debugging common issues, reviewing code for style and security.
What I provide that the agents can't: architectural decisions, client relationships, domain understanding, creative problem-solving, and quality judgment. These are senior-developer activities, and they're the only ones a client should be paying for.
What Clients Actually Care About
In three years of running this, not a single client has cared about my headcount. They care about three things:
- Can you deliver what you promise? Track record matters. Shipped projects matter. Fancy team slides don't.
- How fast can you ship? My typical turnaround is 3-5x faster than agency quotes for comparable scope. Clients notice this immediately.
- Will you be responsive? Solo means one point of contact, zero handoffs, no "let me check with my team." I'm the person who built it, and I'm the person who answers questions about it.
Nobody has ever hired an agency because they wanted more people involved in their project. They hire agencies because they need something built. Headcount is overhead, not value.
The one legitimate concern clients raise is bus factor -- what happens if I'm unavailable? My answer: every project is documented, every system has monitoring, and the codebases are clean enough that any competent developer could take over. The same risk exists with small agencies -- they just hide it behind the illusion of redundancy.
The Infrastructure: My "Team"
Here's the concrete setup that makes solo operation possible:
Compute Infrastructure
- VPS 1 (primary): Development, client projects, MCP servers, n8n workflows. Ubuntu, nginx reverse proxy, certbot SSL. Everything containerized or process-managed.
- VPS 2: Secondary services, monitoring, backup endpoints.
- NAS: Persistent storage, database backups, asset archives.
Total monthly infrastructure cost: about $200. That runs all my development environments, all client staging environments, and all automation services.
AI Agent Layer
Claude Code is my primary development interface. But it's backed by a stack of specialized components:
- 14 MCP servers I configured and run, providing 170+ tools: database access, email, file operations, web search, code search, workflow automation, browser testing, multi-model routing, semantic memory, Cloudflare management, and more. Two are custom-built (Contract Validator, Framework Developer Agent); the rest are community and vendor servers I integrated into my workflow.
- PAL (multi-model orchestration): An open-source tool I configured to route tasks to the best model. Code reviews go to Claude. Large-context analysis goes to Gemini. Fast simple queries go to GPT-4o. Cost optimization built in.
- Specialized agent definitions: Planner, TDD guide, code reviewer, security reviewer, build error resolver, E2E runner, refactor cleaner, documentation updater. Each agent has a specific role and set of instructions.
Automation Layer
- n8n: 20+ active workflows handling deployment notifications, monitoring alerts, scheduled reports, data transformations, and webhook integrations.
- Playwright: Automated browser testing for every client project. Not just smoke tests -- full user flow verification.
- Custom scripts: Database backup rotation, SSL renewal checks, uptime monitoring, log aggregation.
A Typical Day
I wake up and check my monitoring dashboard. n8n has been running overnight workflows: database backups completed, SSL certs are valid, all client sites responding normally. If something failed, I would have gotten an alert.
Morning is client work. I open Claude Code, load the project, and review what was accomplished in the last session. The TASK.md file tracks progress. I pick the next phase, describe the objective, and let the agent work. While it implements, I review the previous phase's output, check the staging deployment, maybe respond to client messages.
Afternoon is typically new project work or infrastructure improvements. Building a new MCP server, optimizing a workflow, deploying a new client environment. The rhythm is: specify, delegate, review, iterate.
I typically do 5-6 hours of focused work. The AI agents do another 3-4 hours of execution during that time. The effective output is equivalent to 8-10 developer-hours per day from a single person.
Why Most Agencies Are Overstaffed
Controversial opinion: at least half the developers at most small agencies exist to compensate for inefficiency, not to produce output. Here's what I mean:
- Junior developers who need mentoring consume senior developer time. Net productivity gain is often negative for the first 6 months. AI agents don't need onboarding.
- Coordination overhead scales quadratically with team size. Stand-ups, code reviews, design discussions, sprint planning. Two developers need 1 communication channel. Five developers need 10. Ten need 45.
- Context switching costs. When a developer works on multiple projects (as most agency devs do), they lose 20-30% of productivity to context switching. My AI agents don't lose context -- each session loads the full project state.
- Inconsistency. Different developers write different code. Style varies, patterns vary, quality varies. My output is consistent because the same system (me + Claude Code + my MCP servers) produces all of it.
I'm not saying agencies shouldn't exist. Large-scale projects with tight deadlines genuinely need parallel human effort. But for the vast majority of projects I see -- web applications, internal tools, API integrations, automation systems -- a single skilled developer with AI tooling can match or exceed a small team's output.
How to Structure a One-Person Operation
If you're considering this path, here's what I've learned about making it sustainable:
Productize Your Services
Don't sell hours. Sell outcomes. "I'll build your inventory management system" is a deliverable. "I'll work 40 hours on your project" is a cost center. Fixed-price, scope-defined engagements align incentives: I'm motivated to ship fast, the client knows what they're paying for.
Build Your Infrastructure Incrementally
I didn't start with 14 MCP servers. I started with zero. Each server was added to solve a specific problem I hit during client work. Most are open-source servers I configured for my infrastructure -- the PostgreSQL MCP server got added because I got tired of writing throwaway query scripts. The Gmail server got integrated because a client needed automated reporting. Two I built from scratch when nothing existed for the use case. Let the work drive the tooling, not the other way around.
Automate Everything That Happens More Than Twice
Deployment should be one command. Database backups should be automatic. SSL renewal should be automatic. Monitoring should alert you, not require you to check. Every manual process you automate frees time for the work that actually generates revenue.
Maintain Boundaries
The risk of solo operation is working all the time because there's nobody to hand off to. I maintain strict boundaries: no client calls after 6pm, no weekend work except genuine emergencies, and a hard cap on active projects (three at a time). The AI agents can work whenever, but I work business hours.
Document Obsessively
Every architectural decision goes in CLAUDE.md. Every deployment procedure goes in a runbook. Every client's infrastructure is mapped in an inventory. Not because I might forget -- because the AI agents need this context to work effectively. Good documentation isn't just for humans anymore; it's input for your AI team.
The Tools That Make It Possible
Quick reference of the specific tools that form the backbone of my operation:
- Claude Code: Primary development interface. Agent loop, MCP support, hooks system.
- n8n: Workflow automation. Self-hosted, no per-execution costs, 400+ integrations.
- Playwright: Browser automation for QA testing. Runs on VPS, no GUI needed.
- nginx + certbot: Reverse proxy and SSL. Simple, reliable, zero cost.
- PostgreSQL: Database for everything. Direct MCP access from the agent.
- Supabase: For projects that need cloud-hosted auth/storage.
- Frida + jadx + Apktool: For mobile reverse engineering work (niche but pays well).
Is This the Future of Software Development?
Not for everyone. There will always be projects that need large teams: operating systems, real-time trading platforms, safety-critical systems. The cognitive load of some domains genuinely requires multiple human minds working in parallel.
But for the bread-and-butter of software development -- web applications, internal tools, API integrations, data pipelines, mobile apps -- the solo-plus-AI model is viable today and will be dominant within five years. The economics are too compelling, and the tooling is improving monthly.
I'm not the only person doing this. I talk to solo developers every week who've made the same transition. The common thread: they were already strong developers who adopted AI tooling early. They didn't replace their skills with AI. They amplified their skills with AI. That's the pattern that works.
If you're running a similar operation or thinking about making the leap, I'm always interested in comparing setups. And if you're a business that needs software built -- fast, lean, with minimal overhead -- that's literally what I do. Get in touch.