What can I actually build with $0?
More than you think. A starter stack that costs nothing.
What can I build with $0?
More than you think. With careful tool choice you can build a real, useful agentic system without spending a cent.
Here’s the actual stack and what it can do.
The $0 stack
LLM: Local with Ollama. Run Llama 3.3, Qwen, or Mistral on your laptop. Zero token costs.
Agent host: Claude Desktop free tier (limited) or open-source Open WebUI on top of Ollama. Both free.
Tool layer: Free MCP servers — filesystem, GitHub (with your own PAT), web search via Brave Search free tier, Slack via free workspace.
Code editor: Cline (free, OSS) on VS Code + your local Ollama as the model. Done.
Browser automation: Playwright MCP server. Open source, runs locally.
Hosting: Cloudflare Pages free tier (500 builds/month, unlimited bandwidth, custom domain).
Storage: Cloudflare R2 free tier (10GB) or just local disk.
Total monthly cost: $0.
What you can actually build
With this stack, real things you can ship:
1. Personal coding agent
- Cline + Ollama (local Llama 3.3 70B) + filesystem MCP + GitHub MCP
- Performance: slower than Claude Code, but free. Good enough for refactors, small features, bug fixes.
2. Research / summarisation agent
- Open WebUI + local model + web search MCP + filesystem MCP
- Drops result into local markdown files
3. Personal automation bot
- Playwright MCP + cron + custom Node script
- Schedule it to do daily tasks: “check this site, save changes, email me”
- Email via the free tier of any SMTP provider (Brevo, Mailjet)
4. Self-hosted PR triage
- GitHub MCP + Slack MCP (or webhook) + local LLM
- Auto-labels, summarises PRs, posts to Slack
5. Documentation chat
- RAG over your own docs (local embeddings via Ollama nomic-embed-text)
- Frontend deployed on Cloudflare Pages
- Vector store: SQLite with sqlite-vec extension, all local
Where free hits limits
Be honest about where $0 stops working:
Quality. Local Llama 3.3 70B is around GPT-4 level for many tasks but worse than Claude Sonnet 4.5 / GPT-5 on hard reasoning. For simple workflows, fine. For complex multi-step agents, you’ll feel it.
Speed. Local LLMs on a Mac M3 do 10-30 tokens/sec. Cloud APIs do 100+. For interactive use, latency adds up.
Browser-Use / Operator-grade automation. Free tools work for simple flows but the production-grade browser agents are paid. Your own Playwright + LLM works for moderate complexity.
Multi-modal. Vision models are catching up locally but still a step behind paid (Claude Sonnet, GPT-5o, Gemini).
Reliability at scale. Running LLMs locally is fine for one user. Hosting them for production traffic gets non-trivial fast.
When to start spending
You’ll know it’s time to spend when:
- You’re hitting rate limits on free tiers regularly
- Latency is breaking UX
- Quality of local model is the bottleneck on results
- You’re spending hours debugging local infra time that paid services would just solve
The first $20/mo (Claude Pro or ChatGPT Plus) is usually a step-change in productivity. After that, marginal returns flatten until you hit $100+/mo.
My recommended starting path
- Week 1: Install Ollama + Llama 3.3 70B + Cline + filesystem MCP. Use it on real work.
- Week 2: Add GitHub MCP + try a code review agent.
- Week 3: Build one workflow — pick a recipe from this site, adapt it.
- Week 4: Decide what to pay for based on what’s actually limiting you.
You’ll have a working agentic system before spending a cent. Then you spend with intent, not on faith.
What to read next
- Tools — the $0 picks — filtered to free tools
- Recipes — easy starts — beginner-friendly workflows
- Why agents still hallucinate — what to expect from local models