Local-first AI
AI software whose data, credentials and execution live on hardware you control, with cloud as an option rather than a requirement. Privacy, cost and ownership follow from that architectural choice.
Ownership is decided by architecture
Where software runs determines who controls the data, regardless of what any policy promises. Local-first AI puts the primary copy — the credentials, the memory, the run history — on hardware you administer, and treats the network as something you use rather than something you depend on for your own data to exist.
Why it matters more for agents than for documents
An agent system accumulates precisely the material you would least like hosted elsewhere: tokens to every connected service, a memory of how your operations work, and a decision log covering everything it did on your behalf. That corpus is far more sensitive than any single document it processed.
The economics change too
Cloud automation platforms meter operations, which quietly caps ambition — you automate what is worth paying per run for. When the runtime is on your own machine the marginal run costs nothing but electricity and whatever inference you choose to buy, and it becomes rational to automate the small annoying things too.
In AGNT
AGNT Community Core runs entirely on your machine — keys, memory and history included — free.
Local-first AI — common questions
Does local-first mean offline?
No. Local-first software uses the network freely — it simply does not require it for your data to exist or your tools to run.
Do I still need an API key for a model?
Only if you want a hosted model. Point AGNT at Ollama or LM Studio and inference happens locally too, at which point nothing leaves the machine.
Can local-first work for a team?
Yes. AGNT pairs local runtimes with optional cloud and self-hosted deployments for shared operations, without moving the default off your hardware.