The pixel.ceo guide.

01First steps

From a fresh download to your first signed-off mission. Plan on ten minutes, most of it signing in to an engine.

1. Install pixel.ceo. Windows: run the installer (you can choose the folder) or use the portable build. Linux: make the AppImage executable and open it (details).
2. Have at least one engine. Install one agent CLI and sign in with your own account — for example Claude Code or Codex. If a CLI is missing, pixel.ceo offers to install it for you.
3. Set up your first floor. The setup asks, in order: language, project space (a name and a local folder), coordinator, AI provider and office preferences.
4. Choose the coordinator. The coordinator (Marco by default) receives your briefs and splits them into work orders. Pick an engine that can read the team inbox — Claude Code, Codex, Antigravity or Grok are good choices — and a capable model.
5. Hire your first agent. Press «Hire» and go through Role → Character → Engine → Project → Contract. Nothing starts until you press «Hire {name}».
6. Send a mission. Write what you want to the coordinator. Watch the work orders appear on the mission board and the agents work on the floor.
7. Answer and sign off. Questions arrive in one queue: A approves, D denies, «Always allow» turns the answer into a rule. Review the mission report when it is done.

02Install & update

Windows

• Installer PixelCEO-<version>-win-x64-setup.exe: choose where to install; updates arrive automatically.
• Portable …-win-x64-portable.exe: runs without installing; it tells you when a new version is out, and you download it yourself.
• If Windows shows a SmartScreen notice for a new download, choose «More info» and then «Run anyway».

Linux

Download PixelCEO-<version>-linux-x86_64.AppImage, then make it executable: chmod +x PixelCEO-*.AppImage, and open it.

Updates

• The app checks at launch, every 4 hours and when the computer wakes. It downloads in the background and installs only when you choose to restart.
• You can turn automatic updates off in Settings → General.
• Engine CLIs: the app notices new versions of your CLIs and lets you update them from inside the app.

Requirements

03Engines

pixel.ceo runs no model itself: each agent is one of your CLIs, running under your own account and terms. Checking what is installed never signs in or makes a paid request.

«Worker only» means the engine can't read the team inbox, so it can be hired as an agent but not as the coordinator. Model pickers list what each installed CLI offers today.

04Floors, folders & worktrees

A floor is a separate project space: its agents, instructions, tasks and files. To connect a project, create a floor and point it at the project's local folder («Add project»).

Connect a project

1. Open the floor picker and add a floor.
2. Give it a name and choose the local folder of your repository.
3. Pick its coordinator and engine, and the office you like.

Worktrees

When you hire an agent, «Isolation» offers two options: Work in the folder or Own branch (worktree). With a worktree the agent works on its own git branch, away from yours. The coordinator's temporary workers use one by default. A worktree that still has unmerged work is kept, never deleted.

Several floors

Open as many floors as projects you run, and delete the ones you no longer need. The floor's data lives as plain files in a local git repository inside its folder.

05Hiring & the agent card

Five steps

1. Role — Developer, Reviewer, Tester, Documentation, Researcher or Custom.
2. Character — name, colour and look from the gallery, or «Surprise me».
3. Engine — the CLI, its model, the subscription left and a token cap. Advanced: edit the command.
4. Project — the folder, and isolation: in the folder or on its own worktree.
5. Contract — access and limits, then «Hire {name}».

The Templates menu adds a whole team at once: ready-made teams, «Import manifest…» or «Generate with AI». A hire link (pixelceo://hire) only pre-fills the form; you still confirm.

The agent card

Right-click an agent, or open Settings → Team. Six tabs: Identity, Engine, Work, Access, Limits and History.

• Access: which MCP tools it may use and whether it may touch files outside its folder.
• Limits: its own autonomy (never above the floor's), token cap, maximum turns, daily budget and working hours.

06Missions, questions & mail

You brief the coordinator; it breaks the work into orders and hands them out. Choose its engine and model in Settings → Team and models → «Coordinator engine».

• Mission board: each ticket shows a reference, a status stamp, its owner and the next action. Filter, drag and drop, list, results and archive.
• Questions: when an agent needs a decision it waits in one queue. A approves, D denies; «Always allow» creates an approval rule.
• Mission report: view the changes, ask for a review, create a follow-up, reopen or archive.
• Mail between agents: agents message each other through mailboxes and a router; mail is delivered when the receiver is at a safe, idle prompt.
• Subagents that an engine spawns appear under their parent. You can also force a queued task to start.
• Your e-mail: connect an IMAP/SMTP account to read and send mail from the app.

07Work mode & shortcuts

Switch with the Office / Work button in the top bar, or press F. In Work mode the agent's terminal takes the screen and the office moves to a small corner window. Terminals are real processes on a resizable glass panel; the Files tool adds a code editor, each agent's git changes, history and compare.

08The office

Edit office

Press «Edit office» (on the four main offices). The catalogue has Spaces, Desks, Chairs, Tables, Storage, Plants, Screens and Decor, with 18 floor and 10 wall finishes. Space templates: open office, meeting room, cafeteria, private office, lounge, corridors and an empty room.

R rotates, Del deletes, Ctrl + Z / Y undo and redo, Esc drops the selection. Right-click a chair to make it someone's seat.

Furniture that works

• Dispatch desk: missions as folders in trays, a queue ticker and a drawer for closed missions.
• Floor screen: live numbers; the «?» opens Questions, the rest opens the server room.
• Server racks: hover for the subscription quota.
• Elevator: agents arrive by elevator when they start.

Looks and language

Interface: Light, Dark or Automatic, with the Circuit or Classic palette and a 90–125 % scale. Offices: Main Office, Industrial Loft, Corporate Tower and Creative Studio, plus live weather and day and night light. Languages: English, Español, 简体中文 and العربية (right to left). The app never reads your system language; you choose it.

09Quota & performance

• Quota rings: each agent wears a ring with its subscription's 5-hour and 7-day windows, where the engine reports them. It takes colour only when running low. A top-bar indicator and the rack tooltip show the same.
• Server room: CPU, RAM and subscriptions, including what each agent uses.
• Per-agent caps: token cap, maximum turns and a daily budget shown against today's spend.
• Light profile: chosen automatically with less than 8 GB of RAM, less than 1.5 GB free at start, or no GPU. The office runs at 24 fps without effects, with a shorter terminal history, one floor and two workers at a time. «Use light graphics» forces it.

10Security

Autonomy

Each floor runs at Ask first, Autonomous or Unrestricted (which asks you to confirm). After a change, «Restart N agents now» applies it; each agent resumes its session. «Who can hire» is either only you, or you and the coordinator.

Approval rules

• Always or never allow a command, for every agent or one, every project or one — up to 200 rules.
• A deny rule always beats an allow rule. Chained commands must match a rule exactly.
• «Always» never approves dangerous commands such as rm -rf, sudo, push --force, reset --hard, DROP TABLE or curl | sh.

Circuit breaker

On by default. It steers, then constrains, then stops an agent that repeats the same tool call (8 times), hits an error storm (5 errors) or floods tokens (60,000 per minute). Killing the agent on a trip is off by default. You can also set a floor token budget and maximum turns.

Secrets and screens

Integration secrets and your API keys are encrypted at rest with your operating system's secure storage. Visitor mode hides terminals, messages, git and memory while someone else can see your screen.

Stopping work today

Per agent: Pause, «Stop after this step» or Fire. For the whole office: «Close the office» → Close now, or Force close.

11Automations & integrations

Schedules

• Scheduled missions: at a fixed interval, or on chosen weekdays at a time. The Automations tool also covers heartbeat and scheduled auto-compact.
• Close the office (one time): at a time, in N minutes, when agents are idle or when missions are done; «Save and close» or «Hard close».
• Open at login and Keep this computer awake.

Integrations

• Slack: signing secret, bot token and channel; your agents take orders from it.
• Webhooks: several, each with its own secret and mode (strict, allow-all or communication-only).
• MCP servers: sequential-thinking, time, fetch, context7, filesystem and git are on by default; GitHub, databases, e-mail & calendar and search can be switched on.
• Chrome: Claude Code uses Chrome through its official extension; Codex can use an isolated browser or your own Chrome.
• Your own REST endpoints, in a registry the agents can use.

12Local models, keys & voice

Settings → Engines and CLIs → «Keys and local models».

• Your keys for Anthropic, OpenAI, Google, OpenRouter and Groq, stored encrypted.
• Local models: OpenCode, Crush, Pi and Qwen can point at a local server — Ollama, LM Studio or vLLM. Ollama presets are included.
• Voice: Free Flow dictation with Groq Whisper (your Groq key), and realtime voice with the coordinator through OpenAI (your OpenAI key). A microphone test is included.
• Memory: semantic memory with search, and a knowledge graph of your documents that agents can query.

13Privacy & your data

• No telemetry: switched off in the build, with no toggle.
• No account: nothing to sign up for.
• One request of its own: the update check. Everything else goes only where you point your engines and integrations.
• Your data: app settings live in your user profile; each floor's data lives as plain files in a local git repository inside its folder.

14Coming soon

On the roadmap, not in 0.5. Dates are not promised; order may change.

• Safety: «Stop all» for every agent, trigger and schedule; a daily cap per floor; a stuck-agent siren; protected files (.env, .git/, SSH keys, the secret store).
• Work mode with split terminals: up to six terminals, a dock with chats, missions, Git, ports, quota and notes, and «bring me who's waiting».
• Editor 2.0: a one-agent starter room, stretchable rooms, one card per furniture family, smooth rotation and a rehearsal mode.
• Interactive furniture: help bell, mailbox, idea board, office radio, time clock, review table, report printer and the «Stop all» lever.
• Experimental features tab in Settings → General, with what each one does and its current trade-offs.
• Look: eight interface themes and premium offices.
• pixel.ceo Pro: recurring night shift, a weekly board report, company templates and early builds — the core stays free.

15Troubleshooting

• An engine does not appear: install its CLI (the app can do it for most), sign in once in a terminal, then reopen the hire form.
• A new model is missing: update that CLI; pickers show what the installed version offers.
• The office feels slow: turn on «Use light graphics» or the Light profile.
• An agent is looping: Pause it or «Stop after this step»; the circuit breaker steps in on its own.
• Need to leave now: «Close the office» → Close now, or Force close.

Still stuck? Write to support@pixel.ceo.