claude-multi
ENIT
Open source soonUnofficial · for Linux

Every Claude you use. One setup.

claude-multi keeps each of your Claude Code and Claude Desktop accounts isolated on one Linux machine, and gives them all the same memory, the same tasks and the same tools.

Get in touchSee how it works →

Needs Linux, Deno and git — made and used on Arch with KDE. The brain is optional and runs on your own server.

Demo data — the console as it looks, with invented sessions and tasks.

One setup for

  • Claude Code
  • Claude Desktop
  • claude.ai
  • the phone
  • the console
What it brings together

Everything a heavy Claude user ends up building — done once, properly.

Each part works on its own; together they make several accounts feel like one assistant that knows you.

Isolated profiles

One folder per account: its own login, settings, skills and even its own Claude Desktop, with its own icon. Work that belongs to one profile never shows up in another.

A console that is live

A local web app: the sessions running now, today's tasks and appointments, the brain, connections and health. Updates are pushed, not polled, and it works offline.

One brain for every Claude

Memory and tasks on your own server, reached over MCP — the protocol Claude uses to call tools — by the apps, claude.ai, Claude Code, Desktop and the phone.

Secrets out of reach

Every credential encrypted in a vault, its key in the system keyring. Never in a config file, a command line or a tool result, and permission rules keep a session away from the key.

Tools, per account

One registry of MCP servers, all off until you turn one on. Gmail, Calendar, Drive, Coolify, Cloudflare and more: each tool knows which account it acts on.

A doctor, and updates that ask nothing

One command, doctor, checks that everything is in place and says how to fix what is not. Claude Code, Desktop and claude-multi update in the background — verified, and one click to roll back.

The brain

A memory that follows you from the terminal to the phone.

The brain is a small service on your server. Every Claude you use connects to it, so what one learns, the others know.

  • ✓Pages that linkMarkdown in seven areas — who you are, projects, clients, people, notes, a diary, an inbox — linked like Obsidian, every version kept.
  • ✓Search by meaningFull text and embeddings, fused; the model runs on the same server, so nothing leaves it to be searched.
  • ✓Tasks everywhereOne list, written from any chat, shown as a board on the web and in the console, with reminders on the desktop.
  • ✓Yours, and only yoursOne database per person, invite-only, a passphrase and a six-digit code from your authenticator app. Backups are encrypted with your own key.
How it works

Three commands, then it takes care of itself.

  1. Your configurationProfiles, accounts and rules live in a folder of yours, apart from the code — keep it in step between machines with git or Syncthing.
  2. InstallOne idempotent command materialises launchers, settings, the console and the tray app. It never deletes: what it replaces is moved aside. --dry-run shows the plan.
  3. Sign in, and checkLog in once per profile. claude-multi doctor verifies every invariant and tells you how to fix what is off.
~
# 1 · your configuration
$ claude-multi init ~/claude-multi-config --name Ann
# 2 · install: launchers, settings, console, tray
$ claude-multi install
# 3 · sign in to each profile, then check
$ claude          # → /login
$ claude-multi doctor
  ✓ profiles: personal, work, research
  ✓ console on http://127.0.0.1:7331
  ✓ secret vault: 6 accounts, every secret here
  ! MCP registry out of sync
      fix: claude-multi mcp sync
  36 pass · 1 warn · 0 fail
Principles

Built the way you would want it built.

Your data stays yours

The brain runs on your server; backups land on your machines, encrypted. claude-multi collects no analytics and sends no telemetry (Claude itself follows Anthropic's terms).

Secrets stay out of sessions

Encrypted in a vault, never in a command line or a tool result; our MCP servers mask what they return.

Nothing behind your back

Sending mail asks first; destructive commands are refused inside a session and handed to you instead.

Plain, and offline

No build step for the console, libraries vendored, bash on the launch path. It starts on a machine that has never been online.

Open source, soon.

The code is being readied for a public release. Want to know more, or try it early? Write to me.

Write to meRead the docs →