Why Local-First Matters

Most software treats the network as the ground it stands on. The cloud is where state lives, where the work is real, where the source of truth sits. Local is a cache: a temporary, disposable copy of something that actually belongs somewhere else.

The workshop runs the other way around. Local is the source of truth. The network is a publishing step.

The inversion

This is a small change in wording and a large change in consequence. When local is primary, the machine in front of you holds the real thing. Work does not stop when the connection drops. Nothing leaves until you decide it should. There is no background process quietly shipping your state somewhere you did not choose.

When the network is primary, the opposite is true on every count. You are renting access to your own work. The connection is a dependency, the provider is a dependency, and the default direction of data is outward, continuous, and hard to see.

Why it holds up

Local-first is usually argued on privacy, and privacy is real: data that never leaves the machine cannot be intercepted, sold, or subpoenaed from a third party. But privacy is the smaller case.

The larger case is control. A local-first system fails in ways you can reason about. The disk, the process, the file: all of it is in front of you. A network-first system fails in ways you cannot see and cannot fix, because the part that broke belongs to someone else. Resilience is not an added feature here. It is what is left when you stop assuming the network is always there.

What it looks like in practice

In the workshop, every project defaults to local. Changes are committed and versioned on the machine, atomically. Nothing reaches a remote until an explicit gate passes: a build, a test run, a security scan. The remote is not a mirror that tracks every keystroke. It is a deliberate, verified export that happens because someone chose to publish, not because the system was syncing in the background.

The models run the same way. Inference, embeddings, transcription: local hardware, local weights, no API round-trip required to develop or run the work. The cloud is available when it is the right tool. It is never the precondition for the work existing.

The limit of the argument

Local-first is not a refusal of the network. A system that can never sync is as broken as one that syncs without consent. The point is direction and intent: local by default, remote on purpose. Syncing is something you choose, not something that happens to you.

The discipline is simple to state and easy to abandon under pressure: local is the source of truth, and the remote is a decision. Hold that line and the rest follows.