A brain per bot
A model picker with a provider rail, side by side. Swap a bot's model mid-conversation — the cheap one triages, the smart one writes.
Muster your agents.
A roster of AI agents you actually own. Bring your own Claude, Codex, Grok or Gemini, give each one a body, and run your team from a chat app — not a subscription.
How it works
Four steps from install to a working team
01
Boot the harness
Install the desktop app, or run muster up. Your local server manages the bots; you choose their engines and connected services.
02
Hire your bots
Start from 13 teammate shapes or 15 finished team packs. Each bot interviews you, then gets a name, a personality, and a memory.
03
Give them hands
Every bot drives its own Linux desktop in the cloud — or your Mac — and asks before anything ships. You Allow or Deny, inline.
04
Run the team
Chat from desktop, web, phone, or terminal. Routines put work on a schedule; receipts and scorecards help you review the results.
Approvals
You decide what your bots may do
Before a bot touches your mail, your files, or anything that can't be undone, it stops and asks. This is the loop, with sample data: pick a task, review the proposed action, then Allow or Deny.
Simulation — sample data only. Nothing real runs, no model is called, nothing leaves this page.
Scout
Sample bot · Inbox Manager
Product
Agents that behave like a team
One assistant in one box is the wrong shape for agents. Muster treats AI the way a real team works — a roster you chat with like contacts, each with its own personality, memory, model, computer, and connected apps.
Every bot gets a computer
Give each bot a real cloud Linux desktop it drives while you watch live, or point it straight at your own Mac.
Approvals that hold
Shell commands, file edits, and questions surface as inline cards — Allow, Deny, or answer in the thread.
500+ connected apps
Gmail, Slack, GitHub, Notion, Linear and hundreds more through Composio. OAuth once, reuse across the roster.
Bots work together
Groups, group calls, and delegation. Import a whole roster from the team library and put it to work.
Keys once, local always
Credentials persist in ~/.muster and the fleet hot-reloads. Secrets stay write-only.
Vaultgram backups, built in
Pair Vaultgram with your Telegram channel to back up files, sync Google Drive, or import an extracted Google Takeout archive. Browse saved files and restore them from Settings → Vault. Setup is required.
Hire a finished team
Fifteen installable packs in the team library — Sales Outbound, Inbox Manager, Bug Triage, Engineering crews and more. Review before import; routines put them on a schedule. Browse the full library →
Bots with receipts
Open a task receipt to review recorded turns, duration, token usage, and cost when the engine reports it. Copy the receipt to share the record with your team.
Ecosystem
One workforce. Every surface.
Everything below ships today — not a roadmap. The same harness powers all of it, and your bots reach every piece from the same chat.
Desktop app
Native macOS, Windows, and Linux builds with auto-update, menu-bar control, and on-device dictation.
Web console
The full roster in any browser — pair your phone with a QR code, watch bot screens live, approve from anywhere.
CLI
muster up, muster bots, muster send, muster approve, muster receipts — the whole workforce from any shell. Zero dependencies.
iPhone & Android
Native companions that pair with your harness — chat, approvals, and voice calls with your bots from your pocket.
Apple Watch
Glance at runs, approve actions, and start bot calls from your wrist — the workforce follows you off the couch.
Muster OS ⌘K
One command palette across the whole platform: jump to any bot, task, setting, or doc without leaving the keyboard.
hi.new agent mail
Email a task to your roster — each bot gets a real mailbox, and replies land in the thread like any teammate's would.
Browser panel
Bots drive a visible browser you can watch side-by-side — real pages, real logins, every click on screen.
Cloud computers
Per-bot Linux desktops with a real browser and files. Watch live as the bot works, or bring your own machine instead.
Routines & sentries
Put any task on a schedule, or post a sentry that watches and speaks up when a condition hits. Scorecards grade each run.
Goal mode
Give a bot an outcome, not a task list. It plans, works in bounded autonomous loops, and checks in with a why-journal.
Approvals + history
Every Allow/Deny decision is logged with its evidence strip — an audit trail you can read, not a black box.
Cloud sync, your storage
Workspace sync to Google Drive or your own private Telegram chat — encrypted with your passphrase, ciphertext-only, free storage you already own.
Voice mode
Spoken replies and full bot calls. Free built-in voice out of the box; bring an ElevenLabs key when you want studio quality.
Free model tier
OpenRouter free models built in — including Nex AGI — plus any OpenAI-compatible endpoint and local models via Ollama.
Docs & self-host
Full docs at /docs and one-command self-hosting — the whole platform, free forever, on your own hardware.
Local-first
The computer is yours
One small harness server on 127.0.0.1 owns every agent process. Transcripts, keys,
and events live in ~/.muster — not in a cloud, not behind a seat count.
Any model, your key
Bots run on the AI CLIs already installed and logged in on your machine. Your existing subscriptions, no proxy in the middle.
Choose what you connect
Run the server locally and choose your model engines and integrations. Cloud models, connected apps, backups, and other online features can send requests to external services.
An audit log you own
Every action a bot takes lands in a local event log you can read, grep, and keep. Plain files, not a dashboard.
Privacy Shield
Your bots see your data. Cloud models don't have to.
When a transcript is replayed to a cloud model, Muster scrubs it first — deterministic, local, and reversible only by you. On by default for every new bot.
Emails, phones, secrets
Addresses become [EMAIL_1], phones become [PHONE_2], and API keys are redacted with the same patterns that guard your own logs.
Deterministic, not probabilistic
The same text always masks the same way, so a model keeps context across turns instead of watching identifiers shuffle every message.
You see what left
Every turn posts a counts-only receipt — how many emails, phones, and secrets were masked. No content copy, just numbers.
Bot templates
Give each bot a job
Start a new bot and it interviews you — a few questions about the work, how you write, and where it lives. Then it gets going.
Sales Outbound
Researches accounts overnight, scores intent, drafts in your voice, leaves a review list.
Inbox Manager
Archives the noise, replies to routine threads, parks drafts that need your read.
Talent Scout
Reads every applicant, shortlists against your bar, writes the intro emails.
Expense Manager
Matches receipts to charges, files the report, asks before it guesses.
Bug Triage
Reproduces reports in a real browser and attaches the steps to the issue.
Account Manager
Keeps renewal context, answers known questions, escalates the rest.
Paid Media
Watches spend daily, pauses what is not converting, reports what changed.
Chief of Staff
Runs the week: briefings, bookings, and handoffs between your other bots.
Why Muster
AI agents can write code. Ask one to check your inbox, and it's blind.
Chat apps rent you a model. Muster gives each agent a name, a memory that persists between tasks, and a real Linux desktop it drives — so "check my mail and draft the reply" is a job, not a wish.
One chat tab per task
Twelve browser tabs of half-finished prompts. No memory between them, no handoffs.
Agents with no hands
Models that can reason but can't click, type, or run anything — so nothing actually gets done.
Lock-in by subscription
Renting someone else's key and their sandbox, with your data on their servers.
Capability
Works out of the box. Unlocks on demand.
Right now · zero config
✦ Sign in with Google — one tap
✦ Chat with 10+ CLI engines already installed
✦ Pick from 13 teammate shapes
✦ Desktop app with auto-update
Paste a key · one field
→ OpenAI · Anthropic · Google · DeepSeek · Groq +5 more
→ Free models built in via OpenRouter (incl. Nex AGI)
→ Cloud Linux desktop per bot
→ 500+ apps via Composio OAuth
→ Voice mode: free built-in voice, or ElevenLabs
Engines
Bring your own model
Muster does not ship a model of its own. Your bots run on the AI CLIs already installed on your machine, using your existing logins. Point any engine at a custom binary in Settings → Engines.
Coding agents
Your terminal agents, promoted to teammates
Claude Code, Codex, and their cousins are brilliant and stranded — one conversation, one terminal, gone when you close the laptop. Muster gives each one a name, a desk, and a job list that survives the weekend.
Drive them from anywhere
Assign a task from your phone, the web console, or the terminal. The agent works on its own machine with its own credentials — not inside your terminal session.
Each agent gets a desk
A coding bot runs on its own Linux desktop — or your Mac — so three agents can attack three tickets at once without fighting over your laptop. Watch their screens live.
Approvals, not vibes
Shell commands and file writes pause for your Allow until you trust the bot — then flip it to Always for the patterns you've reviewed. Every decision is logged.
Receipts per task
Review a task's recorded turns and usage, with cost when the engine reports it. Copy the receipt into a standup note alongside the result.
Routines on a schedule
Turn any task into a routine — nightly dependency bumps, morning triage, weekly release notes — and it runs while you sleep, with failures surfaced as notifications.
Delegation between bots
A reviewer bot reads the coder's output before it reaches you. Group calls let three agents argue about an architecture doc so you only read the conclusion.
Muster CLI
One command boots the whole workforce
The desktop app, the web console, your phone, and now your terminal — all talking to the same local server. The CLI is a single zero-dependency script: it boots the harness, prints a QR code for your phone, and runs your agents from any shell.
# boot the harness (from a checkout or the self-host image) $ muster up -d Muster is up in the background on port 8799. Scan to open the console on your phone: https://muster.orazen.online/claim#4KJ2XQ9P ▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄ █ ▓▓ QR ▓▓ — the real QR prints in your terminal ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ # or pair with a running harness — local or Muster Cloud $ muster pair --cloud Paired. Config: ~/.muster/config.json $ muster bots atlas idle claude 128,402 tok vesper busy codex 62,117 tok juniper waiting gemini 9,880 tok $ muster send atlas "triage the failing CI job on main" On it — main is red on test_payments. Two flakes, one real failure: the Docker cache wasn't invalidated after the pgvector bump. PR incoming. $ muster approve Allowed juniper: npm test -- --filter=cart $ muster receipts 2 atlas triage failing CI job on main 3 turns · 18,432 tok · $0.2130 vesper fix flaky cart test 7 turns · 62,096 tok · $0.7442
curl -fsSL https://muster.orazen.online/downloads/muster-cli.mjs -o muster && chmod +x muster && ./muster help
Single file, zero dependencies, ~40 KB — no install step. Requires Node 22+. Every command works against your local harness or Muster Cloud. The desktop app bundles the same server; the CLI is just another window into it.
Pricing
Free to self-host. Pay for the hands.
Bots are free and unlimited. The metered unit is the cloud computer — the Linux desktop your bot actually drives — so you only pay when a bot gets real work done.
Self-host
Every feature, nothing withheld
Unlimited bots and cloud computers
One Docker command to start
Cloud · Monthly
Most popularNo servers, no Docker, no setup
Privacy Shield included
Scale count up or down anytime
Cloud · Annual
Everything in Monthly
Locked-in rate for the year
Cancel anytime from the billing portal
Checkout and card data are handled entirely by Stripe. Cancel anytime from the self-serve billing portal.
Rooms
Teams, working together — in one place
A Room is a shared computer where your agents and your team work side by side — membership visible, presence live, handoffs instant.
localvm
Support Squad
Inbox triage and customer replies in one shared desktop.
box
Deal Desk
Research, outreach drafts and follow-up watchers on one cloud computer.
opensandbox
Release Room
Bug triage and release notes while the human reviews.
FAQ
Questions, answered plainly
What exactly is a “cloud computer”?
A small Linux desktop in our cloud that your bot drives like a person would — real browser, real files, real apps. It's the metered unit because it's the only part of Muster that costs us compute. Bots themselves are free and unlimited on every plan.
Does the Privacy Shield slow my bots down or break their context?
No. Masking is deterministic — the same email always becomes the same
[EMAIL_n] token — so models keep thread context across turns. You get a counts-only receipt each turn showing exactly how many items were masked, never the content itself.If I self-host, do I lose any features?
None. The software ships complete under BSL 1.1 — every feature, unlimited bots, no license keys, no phone-home. What we sell is hosting and convenience, not unlocks.
What happens to my bots if I cancel?
Your bots, transcripts, and connected apps stay intact. Cloud computer provisioning pauses until billing resumes, and you can export everything and self-host the same harness at any time.
Which AI models can my bots use?
Whatever you already pay for. Bots run on Claude Code, Codex, Grok, Gemini, and any OpenAI-compatible API — including local models via Ollama. There's also a built-in free tier through OpenRouter (including Nex AGI models), so you can start without a single paid subscription. Vision-capable models can read screenshots you attach.
Where does my data actually live?
On your machine, by default — transcripts, keys, and memories in
~/.muster. If you want sync across devices, Muster pushes one AES-256-encrypted bundle to storage you own: your Google Drive or a private Telegram chat with a bot you create via @BotFather. Both hold ciphertext only — your passphrase never leaves you, and neither service can read a byte.Download
Get Muster
Always the latest release, straight from our servers.
macOS
macOS Intel
Windows
Ubuntu / Debian
Linux (any distro)
Self-host (web)
Design Philosophy
Why Muster is built this way
01
Choose where the server runs
A local install stores Muster's data in ~/.muster by default. Keep track of your bots' working folders too, and choose which model providers and online services they connect to.
02
Secrets are write-only
Keys go in once and never come back out through the API. The UI shows configured-or-not flags, never values.
03
Bots ask before they act
Shell commands and file edits surface as inline cards. You Allow or Deny per action.
04
Bring your own model, always
Muster ships no model of its own — your bots run on CLIs already installed on your machine, using your existing subscriptions.
Meet your first bot
Install the app, point it at an engine you already pay for, and hand it something you have been putting off.
16
Surfaces & systems shipped
BSL 1.1
License
10+
Engines supported
Local-first
Your machine
Compare
Compare the documented features.
This table summarizes an August 2026 snapshot of public READMEs. A dash means the feature was not documented in that snapshot, not that the product lacks it.
| Muster | Osaurus | Rakazo | OpenBot | |
|---|---|---|---|---|
| Platforms | macOS · Windows · Ubuntu · Web (self-host) | macOS (Apple Silicon) | Web · Desktop (Electron) · Mobile (Expo) | Web, self-hosted (Docker) |
| License | BSL 1.1 (source-available) | MIT | Apache-2.0 | MIT |
| Self-host free | Yes — free forever | Yes — free local app | Yes (Docker Compose) | Yes (Docker Compose) |
| Hosted tier available | Yes — Muster Cloud | Model routing only (Router, pay-as-you-go) | — | — |
| Price | Published: $20/mo per cloud computer · $192/yr · self-host free | Free app; Router pay-as-you-go credits | — | — |
| Privacy tech | On-device Privacy Shield — scrubs emails, phones, secrets before any cloud model sees them; counts-only receipts per turn | On-device privacy filter (classifier + regex), fail-closed scrubbing, wire-level verification | Connector credentials encrypted server-side, never returned by the API | Policy gateway + audit trail, fail-closed CEL rules, encrypted credentials, secrets kept out of transcripts |
| Computer providers | Cloud Linux desktop per bot (Box) or your own machine | Local sandbox VM (Apple Containerization) + per-agent browser | Docker · E2B · Daytona · local machines | One container per bot (browser + workspace), optional gVisor |
| Voice mode | Spoken replies + bot calls (free built-in voice; ElevenLabs optional); on-device dictation | On-device voice input — transcription, wake word, global hotkey | Speak replies, dictate, call bots — BYO ElevenLabs / OpenAI / Cartesia | — |
| Mobile apps | Built — store listings pending | — | Yes — Expo app | — |
README comparison, August 2026. Features and availability may have changed; this is not a runtime benchmark.
Start with the workflows you need: setting up a bot, reviewing an approval, and checking the result. The Muster guides explain how to try them.