A desk of assistants you talk to. They read, write and organise the files in your projects, talk back in voice mode, and make real Word, Excel and PDF files.

Two Windows apps that run on the AI accounts you already have. Talk to Homedesk Pro about your own files, or hand Homedesk Raptor a goal and get the finished work back, checked.
A desk of assistants you talk to. They read, write and organise the files in your projects, talk back in voice mode, and make real Word, Excel and PDF files.
A team of agents that takes a goal and what success looks like, works in a real browser and your files, and checks the result before it calls anything finished.
Both apps run on your own Windows PC and keep what they make in folders you can open at any time.
Signed in with the Claude Code, OpenAI Codex or Grok accounts you already have, Gemma on your PC, or an API key. When one reaches its limit the work moves to the next.
Each engine runs on your own account and its allowance, and neither app ever signs one out.
By default, sending, paying, posting and deleting wait for your approval, and on a job application one approval covers one send.

A desk of AI assistants for the folders on your PC. Add a folder as a project, pick who to talk to, and ask in your own words, typed or spoken. Made for people who want the help without a terminal.
Each project is a folder on your PC. Reports and documents the assistants make are kept in it, and nothing is ever deleted: what they no longer need goes into a To delete folder for you to clear.
Dictate a message, have replies read aloud, or talk back and forth in voice mode and talk over a reply to interrupt it. The assistants' voices are made on your PC.
Ask for a document, a spreadsheet or a PDF and you get a real file in your project, ready to open in Word, Excel or your PDF reader.
Say "set me up" on the first day and answer five short questions. Each assistant keeps notes about you and reads them at the start of every chat.
The assistants open, read and fill in web pages in Homedesk Pro's own browser. A click that sends, submits, pays, posts or deletes waits for your approval.
Rename the assistants, change their instructions and what each may do, or add your own. Global and project rules set how every one of them works.
Planning

Hand the goal to Raptor and get the finished work back. Say what you want and how you'll judge it. Agents on your own PC plan it, do it in a real browser and your files, and check the result against your criteria.
Raptor is the one you talk to. Ask it how a mission is going and it reads the real record, not a summary; ask it to get something done and it proposes the team, the goal and what success looks like, then waits for you to press Approve. It can't approve its own cards, and it never changes a setting you haven't seen.
It watches every mission without spending a single model call on it, and tells you in its chat, on Windows and on Telegram when one needs you, finishes, or gets stuck.
Give the team a goal and what success looks like. The lead plans the tasks, the others work through them at the same time, and a fresh session that did none of the work checks every criterion against the evidence before anything is called finished.
The agents work in a real browser on your PC, and you can watch any tab live, or sign in to your own accounts in it, such as a job board, so they can use them.
Every mission asks before anything is sent, paid, posted or deleted, unless you switch that off. On a job application the hold sits at the browser itself, and one approval covers one send.
Each criterion gets a tick or a cross with what the check actually found, and a link that doesn't show what its line says isn't counted as met.
Claude Code, OpenAI Codex, Grok, Gemma on your own PC, or your API keys. When one reaches its limit the work moves to the next, and the limited one is tried again within the hour.
Each mission gets a folder in your work home, named by date and title, with a README Raptor keeps up to date: the goal, the team, each criterion with its evidence, and the files the team made. Open it in Explorer at any time.
A mission survives a restart and picks up where it stopped. Set one to repeat every day, every weekday or every week, and it runs at that time with the same team, goal and criteria, with a message when it is ready.
Every change Raptor proposes is a card under its reply that shows every word it will save, with Approve and Decline. Nothing changes until you press Approve.
Even with your approval, Raptor cannot change ask before acting, your work home, your standing instructions, the engines, Telegram or your keys.
Raptor never keeps a job board password. You sign in yourself in the Live Browser Mirror, and its browser remembers the sign-in.
If you link your inbox, Raptor only reads it: nothing is marked read, moved, deleted or sent, and no AI model reads your mail. Replies are matched to your applications by plain rules.
Raptor's server answers only its own page on this PC. Telegram works by Raptor asking for new messages, so no port is opened, and it acts only on messages from the chat that linked the bot.
Raptor's agents do not read your CLAUDE.md, rules, memory, hooks or settings, and are not given the tools that list or message your other Claude Code sessions.
Downloads go out on request for now. Ask for either app, or both, and we'll reply with the installer and how to set it up. Each is one installer for Windows 11 that brings what it needs and installs any engine you choose that is missing.
Yes. They are separate apps and can sit on the same PC, each with its own window, its own folders and its own settings.
The ones you already have: Claude Code, OpenAI Codex and Grok, signed in with your own accounts, Gemma running on your own PC, or API keys you add. You choose which ones are switched on, and Automatic tries them in a set order starting with Claude Code.
Yours. Each engine is signed in with your own account and runs on your allowance, and neither app signs one out. An API key, if you add one, is used last, only when every subscription has reached its limit.
Into folders on your PC. In Homedesk Pro, each project is a folder you choose, with the files the assistants make kept in it. In Homedesk Raptor, each mission gets its own folder in your work home, named by date and title, with a README that Raptor keeps up to date: the goal, the team, each criterion with its evidence, and the files the team made.
Not unless you let it. By default, in both apps a click that sends, submits, pays, posts or deletes waits for your approval. On a job application every send the page makes is held at the browser until you approve it, and an approval covers exactly one send. If you choose to send applications without asking, the application itself goes without approval, but payments, messages, deletions and anything on LinkedIn still ask.
What an engine sends to its own service, as it does when you use that engine yourself; Gemma runs on your PC and sends nothing. In Homedesk Pro, dictation is turned into text by your browser's maker, while the assistants' voices are made on your PC. In Homedesk Raptor, the messages it sends to your Telegram bot, if you link one, and read-only checks of a mail server you choose to connect. Each app's own server answers only its page on your PC.
The app keeps working in the background, with its icon beside the clock. Quit on the icon stops everything.
The mission carries on where it stopped the next time Raptor starts. While a task runs, Raptor asks Windows not to put the PC to sleep; the screen may still turn off, which stops nothing.