Skip to content

Privacy and security

Switchboard runs on your own computers. There is no account to make, no cloud service, and no usage data or crash reports sent anywhere.

What stays home

  • Your chats, files and history stay on your computers and your server. The apps talk only to your server and its hub.
  • The app's log stays on the computer it runs on. It records what the app does, not what your chats say, and your home folder is written as ~. Nothing is sent unless you copy it out yourself (About, Copy diagnostics).
  • Private folders stay private. The file browser hides dot files and private folders, and a file that looks like a key or a secret is listed but never opened: "This looks like a key or a secret, so Switchboard won't open it."
  • New folders need your yes. Before Claude works in a folder it has not been in, Switchboard asks you first, the same question Claude Code would ask.

What leaves your network, and when

What Goes to When
The newest release's list of builds, and the builds GitHub (the server fetches them) An app checks for updates (the server asks GitHub at most every 5 minutes), or downloads one. Updates
A request for the link program and helper GitHub (the new computer fetches them) You add a computer with the app's line while your server has no downloads of its own
iPhone notifications: chat names, machine names and short status lines Apple's push service Only once you add an APNs key to the server. Each kind has its own switch
A waiting chat's last message Claude, on your own plan Only with Summarize waiting chats on
A new chat's first message Claude, on your own plan Only with Name new chats on
A task sent to Auto, and each machine's tools, free memory and repo names Claude, on your own plan Only with Let Claude pick for Auto on
A condensed copy of a chat Claude, on your own plan When you use Start fresh, Continue or Compact into one chat, which write a summary

The three switches are in Settings, AI and your Claude setup, and use Haiku. Summaries use Opus, or Sonnet if you pick it. All of them run Claude Code on your own computers (claude -p), on your own plan, like any other chat.

If you used Switchboard before October 2026

The AI features and the setup changes below are off for a new server. A server that was already running when these settings arrived turned them all on once, so an install that had them kept them. Check Settings, AI and your Claude setup if you'd rather have them off.

Changes to your Claude setup

Also in Settings, AI and your Claude setup, and also off until you turn them on. Each applies to every computer:

Setting What it changes Turning it off
Subagent model rule Adds a short rule to ~/.claude/CLAUDE.md: Haiku for single lookups, Sonnet for search and reading, the default model for design and code Removes only the rule it added
Explore agent on Sonnet Writes ~/.claude/agents/Explore.md, so the Explore agent runs on Sonnet and only reads. One you wrote yourself is left alone Deletes the file only if Switchboard wrote it
Compact at 300k tokens Sets CLAUDE_CODE_AUTO_COMPACT_WINDOW=300000 in ~/.claude/settings.json Removes it only if it's still Switchboard's value

No passwords, on purpose

Switchboard has no passwords. Anyone who can reach its ports (8130 to 8132) can read your chats, type into their terminals and start new chats on your computers. It is meant for your own home network or your own Tailscale network.

Never forward ports 8130 to 8132 to the internet

Use Tailscale (or another private network) to reach Switchboard away from home.

What it does protect:

  • Linking a computer takes a one-time code that works once and expires after 10 minutes.
  • A computer added over ssh gets a key that can only run the Switchboard helper: no shell and no port forwarding.
  • The server answers only to its own names. It checks the Host of each request against the addresses it knows, so a web page you visit can't reach it by renaming itself (DNS rebinding).
  • Updates are signed. The app installs only builds signed with the release key, and checks each file's SHA-256. Why you can trust an update
  • Downloads are checked for damage. install-agent.sh checks sb.py and switchboard-node against the server's SHA256SUMS when it has one, and get-server.sh checks the server kit against the .sha256 beside it (or the one you pass with --sha256). The sums come from the same place as the files, so they catch a broken download, not a server that lies.
  • Nothing runs as an administrator. The server, the helper, the node and the app install for your own account. A few optional steps need one, and you run them yourself: a Windows server's firewall rule, keeping a Linux server's services running after you log out (loginctl enable-linger), and adding a Windows computer over ssh when your account there is an administrator.

Report a security problem

Please do not open a public issue for a security problem. Use GitHub's private report instead: the Security tab of the repository, then Report a vulnerability. Include what you found, how to reproduce it, and which version you ran (About in the app shows it, and Copy version info copies it).

You should hear back within a week. Fixes ship in the next release, and the release notes credit you unless you would rather they didn't. Only the latest release gets security fixes.