Philosophy in practice
Local-first AI tools (and when cloud still wins)
Local-first does not mean “never use a network.” It means the default path works on your machine, with explicit consent for anything that leaves. That default is how we design Continuum, skill gates, and many ClawHub utilities.
Benefits you can feel
- Privacy, fewer accidental secret uploads.
- Cost, hashing a file locally is free; stuffing logs into a paid API is not.
- Reproducibility, stdlib scripts behave the same offline.
- Consent clarity, flags like
--i-consentmake writes intentional.
When cloud is the right tool
Use hosted models for broad knowledge, multimodal leaps, or collaboration you cannot run locally. Be explicit: paste only redacted context; keep source of truth on disk; seal results.
A decision table
| Job | Prefer local | Prefer cloud |
|---|---|---|
| File hash / claim verify | yes | no need |
| Skill risk scan | yes | optional second opinion |
| Creative brainstorm | small models ok | frontier models shine |
| Public search synthesis | limited | often better |
chatagent.ca itself is a static educational site: content is public, tools that touch your files explain that they run in-browser or as local scripts.
What local-first means
Local-first means the model, the file, or the index can live on a machine you control, and the tool still works when the network is slow or absent. It does not mean “no network ever”; it means the network is an option, not the only copy of your work.
The trade-offs
Local tools cost disk, memory, and setup time. They can be slower than a hosted model and harder to update. The benefit is control: you can read the code, keep the data, and decide when to share. A local tool that hides its data path is not local-first; it is a hosted tool with a local icon.
A simple test
Turn off the network and open the tool. If it can still read your files and produce a result, it is local-first for that job. If it shows a login screen, it is not. The test is crude and it catches most of the marketing.
What to keep local
Keep the source files, the index, and anything private on the machine you control. The model can be remote if the inputs are not. The useful question is not “is the model local?” but “where does the data go?”
A practical setup
Start with a small local model for private drafts, a hosted model for public brainstorming, and a shared folder for the files that matter. Write down which tool sees which data. That note is more valuable than a bigger model.
A note on privacy
Local-first is a data-boundary choice, not a personality. A local model can still log a prompt; a hosted model can still be private if the inputs are public. The question is always what leaves the machine, not where the icon lives.
The smallest useful setup
One local model for private drafts, one hosted model for public work, and one folder that holds the files. The setup is small on purpose; a larger stack is harder to audit.
The real test
Unplug the network and try the job. If it still works, the setup is local-first for that job.
