When I started building tools that run entirely on your own hardware, with your own models, no API calls to third parties — I thought I was making a technical choice.

Local inference is faster. No API costs. No latency from network calls. Your data never leaves your machine. These are all good reasons.

Then I started talking about it publicly and realised: the response is never just “oh that’s a cool technical approach.” It’s always more loaded than that.

The moment it became political

I was at a meetup — a tech meetup, nominally apolitical — and I mentioned that all my tools were local-first. Someone asked, genuinely curious: “But why? Are you worried about surveillance?”

Not “is it faster” or “does it save you money.” They went straight to surveillance.

I gave a technical answer. They nodded politely. The conversation moved on. But something had shifted in the room — or maybe in me. I’d said “local-first” and it had been heard as “I don’t trust the cloud companies.”

Which, fine. I don’t. But that’s not why I built it.

The gap between why I build and why people think I build

The actual reasons I built local-first:

  1. Cost. API calls add up. Running Ollama on my own GPU is free after the hardware investment.
  2. Latency. A local API call is ~50ms. A remote API call is ~500ms. For the kind of work I do, that difference matters.
  3. Reliability. I can run my tools on a plane. I can run them with no internet. I don’t want my workflow dependent on someone else’s uptime.
  4. Ownership. If I build something on someone else’s API and they change their pricing, or shut down, or get acquired — my tools break. I don’t want my tools to break.

Those are practical reasons. None of them are “I am making a political statement about big tech.”

But here’s the thing: when you’re building in 2026 and you choose not to depend on OpenAI or Anthropic or Google, people hear the political statement anyway. Because in 2026, that choice IS political. Not because I made it political, but because the context made it political.

The infrastructure question — who hosts your AI, where does your data go, who has access — stopped being a technical question around the time Cambridge Analytica happened. Maybe earlier. What changed isn’t what I built. What’s changed is the world I’m building in.

What this means for the local-first movement

I think local-first AI is having a genuine moment. Not because more people are becoming privacy absolutists, but because:

The economics are finally working. A 4090 has 24GB VRAM. That’s enough to run most open-source models at reasonable speed. The hardware ceiling that made local AI a niche hobby has mostly gone.

The models are good enough. Two years ago, local inference meant running GPT-2 on your laptop. Now you can run Mistral 7B or Llama 3 at speeds that are genuinely usable. The quality gap between local and API has narrowed dramatically.

People are tired of the dependency. API outages make the news now. Model deprecations break production pipelines. The “move fast and ship” approach to AI infrastructure has costs that are becoming visible.

Local-first isn’t a political movement. It’s a risk management strategy that’s starting to make economic sense.

But it still sounds political. Because the alternative — putting everything in the hands of three or four companies that control the infrastructure — IS a political choice. It’s just the default one.

The honest version

I don’t build local-first because I’m worried about surveillance. I build local-first because I want to own my tools.

That said: if my choice to run everything locally makes it slightly harder for any single company to have complete visibility into how a large number of independent developers are building AI tools — that’s not nothing. That’s probably the most politically impactful thing about what I do, and I didn’t even intend it.

The things you build are always more political than you intended. That’s worth remembering.

If you’re interested in local-first AI tooling, agentmemory is on GitHub — MIT licensed, runs entirely locally.