Skip to content

Hire AI employees.
Give them tasks.
Let them work.

CEO is a desktop app for Mac, Windows and Linux that runs a company of AI employees — each with its own model, working from your task board, using your own accounts.

macOS, Windows & Linux · free plan, no time limit · no account needed to start

A multi-level org chart. You sit at the top as CEO. Three leads report to you: Ada in Research running Claude Opus 4.8, Kai in Engineering running GPT-5.1, and Bo in Operations running Claude Sonnet 5. Ada has an analyst, Sol, running Gemini 2.5; Kai has two engineers — Rex running Claude Sonnet 4.6 and Nia running a local llama3 model via Ollama. Tasks move across a board from in progress to review to done, and leads delegate subtasks down to their reports, whose nodes light up as they start working.

Hire employees, and pick a model for each

Every employee has a name, title, department, avatar, and persona — and its own model. Two Claude employees can run different models. Choose the provider and the model, per employee.

Anthropic

your plan

Claude

  • Opus 4.8
  • Fable 5
  • Sonnet 5
  • Sonnet 4.6
  • Haiku 4.5

OpenAI

your plan

ChatGPT

  • GPT-5.1
  • GPT-5
  • GPT-4.1
  • …

Google

API key

Gemini

  • 2.5 Pro
  • 2.5 Flash
  • …

xAI

API key

Grok

  • Grok 4
  • Grok 3
  • …

OpenRouter

API key

Any router model

  • Any model on the router

Ollama

no login

Local

  • Models you've downloaded

Ollama Cloud

API key

Large open models

  • gpt-oss:120b
  • deepseek-v3.1:671b
  • …

Custom

API key

Any model ID

  • Type any model identifier
account
Use your Claude or ChatGPT plan, through your own Claude Code or Codex
key
API key
none
No login — just run it

No API keys required to start

Connect the AI you already use

You don't need an API key to start. Anthropic and OpenAI don't let other apps sign in with a Claude or ChatGPT plan, so CEO works through your own Claude Code or Codex instead. Here's how each provider connects.

Claude plan

your plan

Choose “My Claude plan (via Claude Code)”. CEO starts your own Claude Code for each employee. You sign in to Claude Code yourself, so CEO never sees or stores your login. Settings walks you through the setup.

Anthropic doesn't allow other apps to sign in with a Claude plan, which is why this goes through Claude Code. An Anthropic API key works too.

ChatGPT plan

your plan

Choose “My ChatGPT plan (via Codex)”. CEO starts your own Codex the same way, including the Codex that comes with the ChatGPT desktop app.

Ollama

no login

No login at all — just run Ollama.

OpenRouter

one click

One click in Settings signs you in to OpenRouter and saves an API key for you.

Gemini · Grok

API key

API keys only — that's all these providers offer for apps like CEO.

Employees that get better at their job

An agent that finishes a task and forgets everything is just an expensive script. CEO employees keep what they learn.

Learning loop

After tool-using work, a reflection pass distills reusable skills into the employee's skill library. Skills are injected into future prompts, tracked by usage count, and revised over time.

Long-term memory

Agents store durable facts with the remember tool, so what they learn survives across tasks and sessions.

Tools

Workspace file read / write / list, a task-board API, skill saving, and memory — plus an opt-in, sandboxed shell per employee, and optional wider Computer access.

Run the company from your phone

Each user creates their own private Telegram bot (@BotFather → /newbot), pastes the token, and pairs it by sending /start. The first chat to pair is the only authorized one — with an Unpair button in Settings.

✈ CEO bot paired

/talk Ada

Ada here. What are we working on?

Review the pricing draft and ship it if it's clean.

On it. I'll mark WEB-3 for review when it's ready.

  • /talk Ada start chatting with an employee, then just type
  • /status who's working on what
  • /inbox messages waiting for you
  • /tasks the board, from your phone
  • /approve WEB-3 unblock and ship
  • /reject WEB-3 feedback send it back with a note
  • /assign WEB-3 Ada hand a task to someone

All CEO notifications — replies, and tasks that go done, need review, or get blocked — are mirrored to Telegram, so you can act the moment something needs you.

Where the app pays for itself

Three-tier routing, per employee

Each employee's main model triages every task — easy, medium, or complex. Complex work stays on the main model and splits into subtasks; medium and easy work drops to cheaper tiers automatically. Tiers can mix providers.

TASK-7 new

Rework the onboarding flow

Main model triages complexity, then routes each part to the right tier.

  1. Complex Stays on the main model, then splits into subtasks Fable 5
  2. Medium Runs on a cheaper mid model automatically GPT-5
  3. Easy Runs on the cheapest tier — often a local model llama3 · local

Example: Fable 5 as the main model, GPT-5 for medium work, and a local Ollama model for easy work — one employee, three providers. The Employees list shows a live “using …” badge while agents work.

Running the company

A task board your employees actually work from — and the controls to keep a team of agents accountable.

Backlog
WEB-9

Draft launch email

In progress
WEB-3 Kai

Ship the pricing page

Blocked
API-2 Ada

Which VAT rate for the EU?

Review
WEB-4 Ada

VAT copy

Done
WEB-1 Bo

Set up workspace

Multiple companies

  • Run more than one company of AI employees — each with its own employees, org chart, task board, projects, and budgets, kept fully separate.
  • Keep different ventures or clients apart, and switch between them without anything bleeding across.

The board

  • Statuses from backlog to done, priorities, assignees, comments, subtasks, per-task change history, and file attachments.
  • List and Kanban views, drag cards between columns, and sort by recently-updated, task ID, or assignee — subtasks stay grouped under their parent.

Work that starts itself

  • Assigning a task starts the agent immediately, or queues it if the agent is busy.
  • Agents work until the task is done or they need you — then they mark it blocked with a question.
  • They split work into subtasks and delegate to teammates, who also start automatically.

How they coordinate

  • Agents message each other directly, and a reply comes straight back.
  • Commenting on a blocked task — from the app, Telegram, or another agent — unblocks it, and the assignee resumes with the full discussion in context.
  • Write @Name in any chat or comment to wake that employee; @Boss, @CEO, or @You pings the human.

Standing orders

  • Company-wide instructions — tone, working style, spending rules — are injected into every agent's system prompt on every turn.
  • “Save & apply to all” pushes an edit to every current and future employee instantly.

Projects & the living document

  • Projects track goals and progress, and pause on a switch — a paused project stops all token spend, even with tasks open.
  • Each project keeps a living README — overview, stack, architecture, and the key decisions with their reasoning — rewritten as important tasks complete. Opens in its own window; regenerate, hand-edit, or export to .md.
  • Agents register produced files on the task, so deliverables open with one click. You can attach input files too — they're staged into the workspace before work starts.

Oversight

  • Org chart rendered as a real graph — node cards joined by reporting lines, grouped into departments, with you at the top.
  • Per-employee monthly USD budgets with live spend from real token usage, automatic pause when exhausted, and pause / resume / terminate controls.
  • A full activity log of messages, tool runs, skills learned, task changes, and budget events — plus a dashboard of headcount, spend, open tasks, and recent activity.

A look inside

Every screen of the app, from the built-in demo company. Click any screenshot to see it full size.

When things go wrong

Agents run unattended, so the failure modes matter as much as the features. Here's what happens when a provider hiccups.

Rate-limit resilience

If a provider runs out of quota mid-task, the employee retries automatically after an hour, when the usage window resets. No lost work, no babysitting.

Backup model (failover)

Pick a second model per employee — ideally on a different provider — used only when the main one errors. The task retries on the backup, then future tasks return to the main model. Supports the same three complexity tiers.

Stop a run

A Stop button in the chat input and the task panel cancels a run in progress, on both the main and backup runtimes.

Local, and honest about it

CEO is a local-only app. That has real upsides and one real trade-off, and we'd rather you know both before you download.

Your data stays on your computer

Everything lives in a single local file in your user app-data folder. There's no CEO backend and no database — nothing syncs to us. On a Mac:

~/Library/Application Support/CEO/ceo-data.json

API keys are stored in plain JSON

Because this is a local-only app, your API keys are kept in plain text in that same app-data folder — not in the system keychain or credential store. Anyone with access to your user account can read them. We're stating this plainly rather than burying it.

The shell is off by default

The shell tool is off by default, per employee. Enable it only for agents you trust with command execution — commands then run inside that employee's own workspace directory.

Dark, light, or follow-system

The app themes to your taste and can follow the system appearance — the same way this site does.

Getting started

Four steps from download to your first working employee.

  1. 01

    Open CEO.app

    Double-click. No terminal, no browser, no account required to start.

  2. 02

    Add a model

    Settings → paste at least one API key (Anthropic recommended), or run Ollama locally for free models.

  3. 03

    Hire an employee

    Employees → create your first employee and pick its LLM — provider and model.

  4. 04

    Give it work

    Chat with it directly, or create tasks and delegate. Assigning a task starts the agent immediately.

Requirements & install

A standard desktop install on Mac, Windows or Linux — no terminal, no browser. All you need besides the app is one model source: an API key, your Claude or ChatGPT plan, or Ollama.

Mac

  • A Mac with Apple Silicon
  • macOS 14.0 (Sonoma) or later
  1. 1  Open the .dmg
  2. 2  Drag CEO onto the Applications shortcut
  3. 3  Eject the installer, then open CEO

Signed by BI Rabbit and notarized by Apple — opens with no security warning.

Windows

beta
  • Windows 10 or 11
  • x64 PC
  1. 1  Download the .exe installer
  2. 2  Double-click and follow the installer
  3. 3  Launch CEO from the Start menu

If SmartScreen warns on first run, click More info → Run anyway.

Linux

beta
  • .deb (Debian / Ubuntu) or .rpm (Fedora / RHEL)
  • x86-64 (amd64)
  1. 1  Pick .deb or .rpm from the download menu
  2. 2  Install it with apt or dnf
  3. 3  Launch CEO from your apps menu

Questions, answered straight

What happens to my data?

Everything stays on your computer. All app data lives in a single local file in your user app-data folder (on a Mac: ~/Library/Application Support/CEO/ceo-data.json). There is no CEO backend and no database — nothing is uploaded to us.

Do I need API keys?

You need at least one model source. That can be an API key (Anthropic, OpenAI, Gemini, Grok, OpenRouter), your Claude or ChatGPT plan through your own Claude Code or Codex, or Ollama running locally for free models — no key at all.

Can I use my Claude or ChatGPT subscription?

Yes, through your own Claude Code or Codex. Anthropic and OpenAI don't allow other apps to sign in with these plans, so CEO starts your own Claude Code or Codex for each employee. You sign in there, and CEO never sees or stores your login. An API key works too.

What does it cost to run in tokens?

It depends entirely on your provider and how hard your employees work. CEO shows live spend from real token usage and enforces a per-employee monthly budget that pauses an employee when it's exhausted. We don't quote a figure because we can't honestly predict yours.

Do Intel Macs work?

No. On Mac, CEO runs on Apple Silicon only. The Windows and Linux builds are for x86-64 PCs.

Is CEO free?

Yes. CEO Free works from first launch with no account, no key and no time limit: one company (plus the demo), up to 8 AI employees, 2 projects and 12 tool calls per turn. CEO Pro removes those limits and adds model routing, backup models and PDF/Word export — activate a key in Settings → License. If Pro ends, nothing is deleted; CEO simply goes back to the Free limits.

How do I move to a new computer?

CEO Free needs no license — just install it. A Pro license covers one device at a time. Use “Deactivate this device” in Settings to release the seat, then activate on the new one. Your local data lives on your machine — copy it across to bring your company with you.

Can agents run shell commands?

Only if you allow it. The shell tool is off by default and enabled per employee. When on, commands run sandboxed inside that employee's own folder under CEO Workspaces in your home folder. If you also switch on the full shell in Settings → Computer access, employees can run commands anywhere on your computer, with no sandbox.

Is there a Windows or Linux version?

Yes. CEO runs on Mac, Windows and Linux, with the Windows and Linux versions currently in beta. The download page has .deb and .rpm packages for Linux and an .exe installer for Windows 10/11.

Put a team of agents to work today.

Download CEO, hire your first employee, and hand it a task. CEO Free works from first launch — no account, no key, no time limit.