Here's the uncomfortable truth I stumbled into three months ago: I was spending more on AI subscriptions than on my coffee habit, and getting about the same quality. I think Between GitHub Copilot ($10), ChatGPT Plus ($20), Claude Pro ($20), Cursor ($20), and a couple of niche tools I'd subscribed to during late-night "this will fix everything" impulses... I was out about $150 a month. That's $1,800 a year. For tools that, honestly, weren't making me $1,800 worth of productive.
So I did something drastic. I cancelled everything. And replaced the whole stack with local and open-source alternatives.
This isn't a "Google served me an ad" experiment. This is three months of real production work — shipping features, debugging production outages, and writing documentation — entirely on open-source tools. Here's what worked, what absolutely didn't, and what I'm still paying for.
Why $150/month Felt Reasonable (Until It Didn't)
The thing about AI subscriptions is they sneak up on you. $10 here, $20 there. Each one feels defensible individually. "It's just the cost of doing business in 2026."
But here's what I found when I actually tracked my usage across all of them:
- 57% of my AI interactions were basic code completion — autocomplete-style suggestions
- 22% were chat-based debugging help
- 14% were "explain this codebase" or documentation tasks
- 7% were code review
And the kicker: I was using Claude for the complex stuff, Copilot for inline completions, ChatGPT for quick questions, and Cursor for the IDE integration. Massive overlap. I was essentially paying four different companies to do slightly different versions of the same thing.
The Open-Source Stack That Replaced It All
I spent the first week just experimenting. Here's what survived the cut:
1. Code Completion: Continue + Ollama
Continue with a local model via Ollama replaced Copilot and Cursor's inline suggestions. I'm running Llama 3.1 8B quantized to Q4_K_M — it fits in 6GB of VRAM and runs fast enough for tab completion.
The trade-off? It's not as good as Copilot at guessing multi-line refactors. But for single-line completions — which is what I actually use 80% of the time — it's within spitting distance. And it costs exactly $0.
Setup time: 20 minutes. If you already have Ollama, it's literally: ollama pull llama3.1:8b and install the Continue VS Code extension.
2. Chat & Debugging: Claude Code (free tier) + Goose
For the heavy lifting — debugging, code review, refactoring — I use Claude Code on the free tier when I need Anthropic-level reasoning, and Goose (50,725⭐ on GitHub) for everything else.
Goose is the dark horse here. It's open-source, extensible, and unlike a lot of AI coding tools, it actually executes code rather than just suggesting it. I had it debug a race condition in a Node.js microservice last week — it added logging, ran the tests, identified the culprit, and proposed a fix without me touching the keyboard.
3. Documentation: OpenWiki
OpenWiki (8,093⭐) is a CLI tool that writes and maintains agent documentation for you. I pointed it at my project and it generated a surprisingly solid CLAUDE.md and contributing guide. The docs aren't perfect — I edit them — but having a first draft saves me hours.
4. Local LLM Inference: whichllm + RamaLama
When I need to run models locally, whichllm (5,596⭐) benchmarks which model actually performs well on my specific hardware. And RamaLama (2,948⭐) handles the serving side — pulling models from any source and making them available via a local API.
This combo replaced my entire cloud API dependence for routine coding queries. I'm hitting Ollama maybe 200 times a day now. Zero API costs.
What I Actually Lost
I'm not going to pretend the open-source stack is strictly better. Here's what I genuinely miss:
GitHub Copilot's context awareness. Despite all the criticism, Copilot is really good at understanding your project structure and suggesting multi-file edits. Continue + a local model just... doesn't do that yet.
Cursor's agent mode. Cursor's Composer/Agent mode that can edit multiple files, run terminal commands, and handle the whole "build this feature" workflow is genuinely impressive. The open-source alternatives (like Goose, Aider) are getting there, but they're not as seamless.
Speed. Local models are slower. Llama 3.1 8B gives me maybe 30-40 tokens/second on my RTX 3060. Compare that to GPT-4o's near-instant responses. You notice the difference.
The Numbers That Surprised Me
After three months, here's what actually changed:
| Metric | Before (Paid) | After (Open Source) |
|---|---|---|
| Monthly cost | $150 | $0 |
| Features shipped | ~8/month | ~10/month |
| Bug fix time | ~2h avg | ~1.5h avg |
| Code review time | ~45min | ~35min |
| Frustration with tool | Medium | Low |
| Setup complexity | None | Mod (first week) |
The feature count surprised me most. I shipped more after switching, not less. I think it's because I stopped context-switching between four different AI tools and just... used one or two that I knew well.
Where Open-Source AI Fails (Be Honest)
I tried replacing everything, and some things just don't work yet:
- Design/code generation from screenshots. No local model can competently take a mockup and generate a React component. I still use Claude's free tier for this once a week.
- Long context windows. Local models with 128K context are still rough — they lose track of details past ~30K tokens. For large codebase analysis, cloud models win.
- Multi-modal. Want to ask an AI about a screenshot of a bug? Local models either can't do vision or do it badly.
But for 80% of what I do — writing code, debugging, reviewing PRs, generating docs — the open-source stack handles it fine.
The Bottom Line
I'm not saying cancel all your subscriptions tomorrow. The paid tools are genuinely better at certain things.
But if you're like me — paying for 3-4 overlapping AI services and not really tracking whether they pay for themselves — try a month on the open-source stack. Keep one paid account as backup. See what happens.
For me, the savings ($1,800/year) combined with actually shipping more code made it a no-brainer. The tech is good enough now. And it's only getting better.
What's your AI stack look like? I'm genuinely curious — are you all-in on one tool, mixing a few, or have you gone full open-source? Drop your setup in the comments.
Top comments (0)