Run multiple Claude Code / Codex / Gemini / Antigravity subscriptions side by side: one shared brain — skills, agents, memory, settings — separate logins, and every subscription's live 5h/7d limits in a single view.
You have multiple Claude Code subscriptions (personal, work, client) each in separate ~/.claude-* directories. You maintain symlinks manually, duplicate settings, and juggle bash functions to switch between them.
aimux treats your AI CLI configs like tmux treats terminals: one shared brain, multiple isolated sessions.
npm install -g @digital-threads/aimux
~/.claude + extra directories (~/.claude-work, etc.)npm install -g aimux
aimux init # auto-detects all ~/.claude* dirs
aimux status # verify profiles, auth, symlinks
aimux run w # launch work profile (prefix matching)
~/.claude (one subscription)npm install -g aimux
aimux init # creates config with main profile
aimux profile add work # add a new profile
aimux auth login work # OAuth for the new account
aimux profile update w -m 'opus[1m]'
aimux run w
aimux profile add myapi --api
# Configure API endpoint (leave blank to use default):
# Base URL: https://api.your-provider.com/v1
# Auth token: [hidden]
# Default model [claude-sonnet-5-5]:
# Opus model [claude-opus-5-5]:
# Sonnet model [claude-sonnet-5-5]:
# Haiku model [claude-haiku-4-5]:
# ✓ Credentials saved to ~/.aimux/profiles/myapi/.env (chmod 600)
aimux run myapi
See Per-profile environment variables for the declarative alternative (.env file / env: block) used by power users and CI.
# Install Claude CLI first, then:
claude auth login # creates ~/.claude
npm install -g @digital-threads/aimux
aimux init
aimux profile add work
aimux auth login work
aimux run # interactive picker (↑↓ + Enter)
aimux run w # prefix match → work
aimux run o -m claude-sonnet-5-5 # one-time model override
aimux run w --resume # flags pass through to Claude CLI
aimux run --auto # launch whichever subscription has the most headroom left
aimux status # dashboard
aimux usage # token usage by profile for the last 7 days
aimux usage --all # all known transcript usage
aimux memory # Claude's auto-memory for this folder: notes saved twice, notes MEMORY.md misses
aimux memory export # copy those notes into the repo (.claude/memory), never overwriting one you changed
# Set a default model per profile. Prefer a family alias: `opus[1m]` always means the
# newest Opus, so a new release is picked up without touching aimux. A full id such
# as `claude-opus-5-5` pins that exact version until you change it. Quote names with
# brackets, or the shell will try to expand them.
aimux profile update w -m 'opus[1m]'
aimux profile update w -m claude-opus-5-5 # pin one version
aimux profile update w --unset-model # let the CLI choose its own default
# Set a fallback model, tried automatically when the primary is overloaded/unavailable
aimux profile update w --fallback-model claude-sonnet-5-5
aimux profile update w --unset-fallback-model # remove it
aimux use)aimux run launches a one-off session. If you'd rather activate a profile so
plain claude / codex use it — like nvm use or pyenv shell — enable the
shell integration once:
# add to ~/.zshrc or ~/.bashrc (fish: ~/.config/fish/config.fish)
eval "$(aimux shell-init)"
Then:
aimux use work # activate 'work' in this shell (persistent until you switch)
claude # runs under 'work', on its model — no `aimux run` needed
codex # same; the CLI adapter sets CODEX_HOME for you
aimux use api # switch profiles — stale ANTHROPIC_*/tokens are cleaned up
aimux use # no name → interactive picker
Each shell is independent, so different terminals can hold different active
profiles at once. The switch only exports env vars into the current shell — it
never changes global state. aimux run still works for one-off launches, and it
always runs the profile you name, whichever one the shell has active.
aimux split)aimux split # one tmux pane per logged-in claude subscription, in this directory
aimux split work client # just these two
Each pane runs a full aimux run <profile>, labelled with its profile on the pane
border; a pane whose run fails stays open with its error. Closing the window ends the
panes, as in any terminal — the conversations are saved, and aimux run <profile> --resume picks one up again. Inside tmux it opens a new window instead of nesting
tmux, and that window is yours to keep or close. Needs tmux (sudo apt install tmux,
or brew install tmux).
Transcripts are shared, so a session can carry on under another subscription — only
the login changes. When a claude session started with aimux run stops because its
window is spent, aimux offers to continue it where there is the most room left:
⚠ work hit its 5-hour limit (resets 23:10).
Continue this session on personal (5h 10%, 7d 16%)? [Y/n]
Press Enter and the same session resumes there, with the flags you started it with.
aimux reads the hit from the session's own transcript, so an ordinary exit costs
nothing; the other subscriptions are only checked when a limit was actually reached. It
works inside aimux split panes too.
A claude session started with aimux run — or in an aimux split pane — shows every
subscription in a band above the prompt, the one it runs on first:
/aimux, aimux status and the dashboard say when it is a week away, and
aimux auth login <profile> once it has passed./exit is already in the prompt:
press Enter and aimux carries the conversation over to that subscription without
asking again./aimux prints the whole table, each window with its reset time.It is a Claude Code mod (Claude Code 2.1.287 or newer), loaded only into the sessions
aimux starts; AIMUX_NO_MOD=1 turns it off. The same figures are available to scripts
as aimux status --json.
aimux dashboard)aimux dashboard
opens a page with every subscription in one place:

~/.aimux/limits-history.jsonl, a megabyte at most), and each card draws them.The page speaks English and Russian — it follows the browser, with a switch in the corner — and has a light and a dark theme.
aimux serves the page itself, on http://localhost:4747 and to this machine only. The
address it prints works until you stop it, and the figures answer only to a page opened
through it: not to another program, another user of the machine, or another web page
open in the browser. It reads the transcripts on this machine — the ones aimux usage
reads — and asks each provider for the limits the way aimux status does. --port <n> picks another port,
--no-open prints the address without opening a browser. Dollar figures are list-price
estimates, what the same tokens would cost through the API: good for comparing
subscriptions, not what a subscription bills.
aimux mcp)aimux mcp is a read-only MCP server over stdio. It has three tools: aimux_limits
(the limits with the verdicts on the pace), aimux_usage (where the tokens went in a
window) and aimux_limits_history. To add it to Claude Code:
claude mcp add aimux -- aimux mcp
| Command | Description |
|---|---|
aimux init | Auto-detect Claude dirs, create config, migrate profiles |
aimux init --source <path> | Initialize with explicit source directory |
aimux status | TUI dashboard — profiles, auth, live 5h/7d limit usage and reset times (claude + codex), auto-mode rules when any are set, symlink health. --watch [s] keeps it on screen and re-reads the limits; --json prints them for scripts; --max-age <s> reuses a reading that recent |
aimux status --no-limits | Same dashboard without the rate-limit probe (offline / faster) |
aimux usage | Show token usage by profile (Claude transcripts + codex rollouts), including sessions started outside aimux. A └ background line under a profile is the part its subagents and plugins' own claude runs spent |
aimux usage --profile work --since 24h | Show usage for one profile over a recent window. --json prints the full report: per profile, day, model, project and the heaviest sessions |
aimux memory | Claude's auto-memory notes for the current folder (shared by every profile): notes saved twice, notes MEMORY.md doesn't list, index lines pointing at nothing. --all lists every project with notes; aimux memory export [dir] copies them into the repo (default .claude/memory), leaving any file you changed there alone |
aimux dashboard | Open a page in the browser: every subscription's limits and pace, and where the tokens went. --port <n>, --no-open |
aimux mcp | Serve the same figures to an AI client over MCP (stdio): claude mcp add aimux -- aimux mcp |
aimux run [profile] | Launch AI CLI with correct env and model |
aimux run | Interactive picker — history pre-selects last used profile |
aimux run w | Prefix matching — launches work if unambiguous |
aimux run work -m claude-sonnet-5-5 | Launch with model override |
aimux run --auto | Probe every subscription's live limits and launch the one with the most headroom (stays within the same CLI) |
aimux split [profiles...] | Open several subscriptions side by side, one tmux pane each (default: every logged-in claude subscription) |
aimux use [profile] | Switch the current shell to a profile (persistent) — plain claude/codex then use it. Requires eval "$(aimux shell-init)" in your rc |
aimux shell-init | Print the shell function that enables aimux use (add to ~/.zshrc/~/.bashrc/fish config) |
aimux agents | Multi-profile agent view — see and manage claude background sessions across all profiles in one TUI |
aimux profile add <name> | Create new profile with symlinks |
aimux profile add <name> --api | Create a 3rd-party API profile (interactive endpoint + token prompt) |
aimux profile add <name> --cli codex | Create a profile for another AI CLI (e.g. Codex) |
aimux handoff <sessionId> --to <profile> | Continue a session under another profile/CLI via summary handoff |
aimux profile update <name> | Update model/cli settings |
aimux profile update <name> --fallback-model <model> | Set a fallback model, used when the primary is overloaded/unavailable |
aimux profile update <name> --unset-model | Remove the default model, so the CLI uses its own (newest) default |
aimux profile update <name> --unset-fallback-model | Remove the fallback model |
aimux profile update <name> -e KEY=VALUE | Set an env var in the profile .env file |
aimux profile update <name> --unset-env KEY | Remove an env var from the profile .env file |
aimux profile list | List all profiles (same table as aimux status; --no-limits skips the probe) |
aimux profile remove <name> | Remove profile and clean up |
aimux profile clone <src> <name> | Clone profile with private files |
aimux rebuild [profile] | Sync symlinks and surface local shared-file conflicts |
aimux doctor | Health check — broken symlinks, missing shared entries, conflicts |
aimux auth login <profile> | Launch OAuth flow for a profile |
aimux auth status | Show auth file status per profile |
aimux setup-shell | Auto-install shell completions (bash/zsh/fish) |
aimux migrate isolate | One-time migration: convert per-profile jobs/, daemon/, projects/ symlinks into real private dirs so each profile gets its own supervisor and sessions. Safe — no data is deleted. Add --dry-run to preview. |
All profile commands support prefix matching: aimux run w → work, aimux profile update o → own.
~/.claude/ ← source of truth (your main profile)
agents/
skills/
commands/
memory/
settings.json
.credentials.json ← private, stays here
~/.aimux/
config.yaml ← aimux config
profiles/
work/
agents/ → ~/.claude/agents ← symlink (shared)
skills/ → ~/.claude/skills ← symlink (shared)
memory/ → ~/.claude/memory ← symlink (shared)
plugins/ ← real dir (shared content, per-profile metadata)
marketplaces/ → ~/.claude/plugins/marketplaces ← symlink (shared)
cache/ → ~/.claude/plugins/cache ← symlink (shared)
known_marketplaces.json ← real file (paths point inside this profile)
installed_plugins.json ← real file (paths point inside this profile)
.credentials.json ← real file (private)
.claude.json ← real file (private)
own/
...same pattern...
When you run aimux run work, it sets CLAUDE_CONFIG_DIR=~/.aimux/profiles/work and launches the CLI. Claude sees a complete config directory — shared content via symlinks, private auth locally.
Plugins are shared too, but Claude validates that a marketplace's installLocation
lives inside the active config directory. So each profile gets a real plugins/
directory: the heavy content (marketplaces/, cache/) is symlinked to the shared
~/.claude/plugins, while known_marketplaces.json and installed_plugins.json are
real, path-rewritten copies. ~/.claude stays the source of truth — install or update
plugins from your main profile (or with CLAUDE_CONFIG_DIR=~/.claude claude plugin …)
and every profile picks them up on its next run. A plugin installed from inside a
profile is merged back into the shared source automatically.
aimux isn't claude-only. A profile's cli field selects which AI CLI it runs, so you
can keep claude, Codex, Gemini and Antigravity subscriptions side by side — same shared brain,
isolated auth — and even hand a live conversation from one to the other when a limit hits.
aimux profile add codework --cli codex # a Codex profile
aimux profile add gem --cli gemini # a Gemini profile (~/.gemini)
aimux profile add ag --cli agy # an Antigravity CLI profile (~/.gemini/antigravity-cli)
aimux auth login codework # runs `codex login` under an isolated CODEX_HOME
aimux run codework # launches Codex with the right model flag
aimux agents # claude + codex sessions in one view (CLI-badged)
CLAUDE_CONFIG_DIR /
CODEX_HOME / GEMINI_CLI_HOME), so subscriptions never collide.shared_sources maps each CLI to its source
(claude → ~/.claude, codex → ~/.codex); the legacy shared_source stays as the
claude alias.private; Codex shares a
knowledge allowlist (skills, rules, memories) and keeps auth.json /
config.toml / sessions/ private; Gemini shares GEMINI.md / skills / commands
/ extensions / memories and keeps auth + settings.json + history private.GEMINI_CLI_HOME at the profile's parent and the profile dir IS gemini's .gemini.
Gemini resumes by per-project index (-r latest), not by session id, so cross-Gemini
native resume isn't wired into the session list yet.agy reads ~/.gemini/antigravity-cli from the home and has
no variable to move it, so an agy profile runs with a stand-in home (HOME, or
USERPROFILE on Windows): the profile's own dir. Every sync links the rest of your
real home into it, so the tools the agent runs still find .gitconfig, .ssh and the
rest, and skills, plugins and MCP servers in ~/.gemini/config are shared. Settings and
conversations stay private. The login does only where agy keeps it in a file, as it
does with no system keychain (WSL, a headless Linux). On Windows, macOS or a Linux
desktop agy keeps it in the keychain (Credential Manager, Keychain, Secret Service),
one for the whole user: every agy profile then shares your usual login, and signing in
or out in one does it for all; aimux auth login says so when it happens. aimux auth login ag starts agy, which asks you to sign in. An agy profile can't be switched
into a shell with aimux use: start it with aimux run ag.Hit a subscription limit mid-session? Continue it under another CLI:
aimux handoff <sessionId> --to codework
The two CLIs' transcripts are mutually unreadable, so this is a summary handoff, not
a native resume: aimux reads the source transcript, summarizes it with the target
profile (self-contained — the target CLI does the summarizing), then launches the target
seeded with that summary. Same-CLI continuation (claude↔claude, codex↔codex) still uses
native resume (aimux run <profile> --resume <id>). Note: the conversation context
carries over, not the model — the target continues with its own model.
Many providers expose an Anthropic-compatible endpoint, so they run on the claude CLI
itself — just a different base URL + token. A provider profile is therefore a claude
profile: it shares the full claude brain (skills, plugins, settings, memory, transcripts)
and you can even --resume a claude session under it natively (same CLI, same model swap).
One command, prompts only for the token:
aimux profile add ds --provider deepseek # fills base URL + model mapping
aimux run ds
Built-in presets: deepseek, kimi, glm, qwen, minimax, mimo. Base URLs are verified;
model names drift — update with aimux profile update <name> -e ANTHROPIC_MODEL=….
For anything else (local models via Ollama/LM Studio, a proxy, Bedrock/Vertex), use the
generic aimux profile add <name> --api and point ANTHROPIC_BASE_URL at it (see below).
Costs shown by aimux usage are list-price estimates (Claude, codex/gpt-5, and the
Anthropic-compatible providers) and will drift as prices change.
Some Claude Code modes (3rd-party proxies, self-hosted gateways, Bedrock, Vertex) are activated by environment variables rather than OAuth. aimux injects per-profile env into the spawned claude process (and into aimux auth login <profile>) from two sources, merged in this order:
<profile>/.env — a dotenv file inside the profile directory. Best for secrets. Written with chmod 600 when aimux creates it; aimux run warns if it becomes group/other-readable.env: block under the profile in config.yaml — best for non-secret toggles you want versioned. Overrides .env on key conflict.The fastest way to set up an API profile is the interactive prompt:
aimux profile add myapi --api # prompts for Base URL, hidden token, models
aimux profile update myapi -e ANTHROPIC_MODEL=claude-opus-5-5 # edit later
…which writes something like:
# ~/.aimux/profiles/myapi/.env — do not commit
ANTHROPIC_BASE_URL=https://api.your-provider.com/v1
ANTHROPIC_AUTH_TOKEN=sk-your-token...
ANTHROPIC_MODEL=claude-sonnet-5-5
ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5-5
ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5-5
ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5
The .env parser supports KEY=value, export KEY=value, comments, and single/double-quoted values (with \n/\t escapes inside double quotes). It does not do ${VAR} interpolation or multi-line values — it's a secrets loader, not a full dotenv-expand. .env is always private (never symlinked to the shared source).
# ~/.aimux/config.yaml
version: 1
shared_source: /home/user/.claude
profiles:
main:
cli: claude
path: /home/user/.claude
is_source: true
work:
cli: claude
model: opus[1m] # follows the newest Opus
path: /home/user/.aimux/profiles/work
myapi:
cli: claude
model: claude-sonnet-5-5
path: /home/user/.aimux/profiles/myapi # secrets live in this dir's .env
# Optional non-secret env injected into the spawned CLI (overrides .env).
# env:
# ANTHROPIC_DEFAULT_OPUS_MODEL: claude-opus-5-5
private:
- .credentials.json
- .env # API credentials — never symlinked, never committed
- .claude.json
- policy-limits.json
- mcp-needs-auth-cache.json
- remote-settings.json
- settings.local.json
- stats-cache.json
- statsig
- telemetry
- state # MCP protocol verdicts, per install
- session-env # per-session environment
- backups # copies of the private .claude.json
- remote # Remote Control daemon + its binaries
- security # Agent SDK venv and logs
- shell-snapshots
- cache # fetched model catalog, per account
The defaults above are always applied — listing extra entries adds to them. Claude
Code keeps growing new runtime directories, so if aimux doctor reports a conflict
on one, add it here and run aimux rebuild.
Third-party tools that read aimux profiles:
MIT. The dashboard page ships the Geist and Geist Mono typefaces, © The Geist Project
Authors, under the SIL Open Font License 1.1 (dashboard/fonts/).
No reviews yet. Be the first to rate this tool.
Sign in to leave a review.