Building a local-first AI that asks before it acts
Most AI assistants are a single model behind a chat box. Paragon is something different: an AI orchestration system that runs on its own machine, plans each task with a local model, brings in cloud AI only when a task needs it, and stops for a person when a choice is theirs to make. This post explains the ideas behind it and why we think they matter for trustworthy AI.
Local-first: the thinking starts at home
Every request reaches a local model first. Ours is called Bonsai, and it runs on an ordinary graphics card. Bonsai reads the request and writes a short task brief before anything leaves the machine. That brief decides the route: some work stays local, and some goes to a cloud model.
A local-first AI keeps private work private by default, and it means the system still has a plan when a cloud service is slow or unavailable.
Cloud AIs advise; Paragon decides
When a task needs heavy lifting, Paragon assigns it to a cloud model such as Claude or ChatGPT, which we call a Director for that task. Directors work inside limits Paragon sets: how many calls, which tools, how long. They can suggest anything. They can't raise their own permissions or approve their own work.
Being able to do something isn't permission to do it
This is the core of Paragon's design. A permission layer called Jack10 answers one question, "is this allowed?", and no model can change its rules. A model might be able to draft and send an email; sending it to a client still needs a person's approval. Keeping capability and permission apart is how AI agent permissions stay meaningful as the agents get more capable.
Memory that survives a restart
Paragon writes things down before they count. Tasks, decisions and results are saved, and every day the system takes a backup and checks that it can actually be restored. AI memory is only useful if it's reliable, and reliability has to be tested, not assumed.
Claims need evidence, and "unknown" is an answer
The Truth Protocol checks results against their sources. A claim can be supported, contradicted, or simply not in the evidence. That third answer matters: a system that's allowed to say "I don't know" makes far fewer confident mistakes. You can help calibrate it with the short reference tests on our site, which give Paragon's AI fact checking a human baseline.
Work while you're away, with a person in the loop
Dream State is Paragon's background mode. Its passive form, DayDream, handles upkeep while the machine is idle: tests, backups, integrity checks and tidying up, plus experiments that only ever produce proposals. Longer autonomous work needs a person to approve its scope and limits first. That's what human-in-the-loop AI means to us: people make the decisions that matter, and the system makes those decisions easy to review.
Want to see it? Play a Dream State run: you make the calls at each fork, and ChessMaster forecasts where each choice leads.
Where Paragon is now
Paragon isn't publicly available yet. It's built in tested checkpoints, and each one lists what it doesn't cover as well as what it does. Failures stay on record next to the passes. The next checkpoint asks whether Paragon can carry a full day of dependent work while things break on purpose, and finish with nothing lost or done twice.