Self-hosted Claude: what runs where

You cannot run Claude itself on a machine you own. You can own everything around it. What runs on your mineral, what goes to Anthropic, and the honest alternative.

You cannot run Claude itself on your own server. Nobody can, outside Anthropic and the large cloud platforms that offer it. What you can own is everything around it: the machine, the memory, the keys, the schedule and the connections. That is what Crads-AI is, and this page draws the line exactly so you can decide whether the split suits you.

Claude itself only runs at Anthropic#

Anthropic does not release Claude's weights. There is no download, no container, no version you can put on a server of your own. So a product that says it "self-hosts Claude" is doing one of two things: running the software around Claude on your machine and calling Anthropic for the thinking, or running a different, open model under Claude's name. Crads-AI is the first kind.

How your assistant thinks has the picture behind this: a mind you rent, a memory you own, senses you add. Self-hosting moves the memory and the senses onto your machine. The mind stays rented.

What runs on your mineral#

Your mineral is a server in your own Hetzner or DigitalOcean account, opened only by your own SSH key. On it:

  • Your brain: every page your assistant knows about you, as markdown files in git, readable in any text editor.
  • Claude Code, the program your assistant works in. It runs on your mineral, reads your files there, and calls Anthropic when it needs Claude to think.
  • Your credentials: your Claude sign-in, and the token for every service you connect.
  • Your schedule: the morning brief, the inbox sweep, every job that runs while you sleep.
  • Your run history: what it did, and when.

Each night the brain is backed up to a private GitHub repository in your own account. None of it is copied to Crads-AI, because there is nowhere of ours for it to go.

What goes to Anthropic, every time#

The thinking happens at Anthropic, so whatever the thinking needs goes there:

  • what you type to your assistant
  • the pages and files it reads to answer you
  • whatever comes back from the tools it uses along the way, such as an email it fetched to summarise

That traffic runs on your own Claude account, under your own agreement with Anthropic. Crads-AI is not in the path and never sees it. What Anthropic keeps, and whether it can be used to improve their models, is set by their terms and by the privacy settings in your Claude account. Check those yourself.

Two connections live with Anthropic, not on your mineral#

Shopify and Xero only let Claude's own connector in, so you connect them to your Claude account instead. They work in your chats and not in your scheduled jobs; Shopify and Xero has the detail. Every other connection signs in on your mineral and keeps its token there.

On this computer moves the machine, not the mind#

The app's On this computer mode makes your mineral a folder on your own laptop instead of a rented server. Your brain and credentials live there. Every answer still comes from Claude, at Anthropic, on your account. What you give up is everything that needs a machine that stays on: no scheduled jobs, no Telegram, no connections.

If nothing may leave your building, use a local model#

Some people need every word to stay on hardware they own: a clinic under strict rules, a firm whose clients forbid it. For them self-hosted Claude does not exist, and the honest answer is an open model they run themselves, with a tool like Ollama and a chat window on top. They trade some quality for complete locality, and for that requirement it is the right trade.

I built Crads-AI on Claude anyway. For an assistant that drafts in your voice and reads a week of your inbox, the quality of the thinking is most of what you notice. The part that compounds over months is the memory, and that part is entirely yours.

If that split suits you, your first hour goes from download to an assistant that is awake.