Deskly is an open-source, first-person desktop app built with Electron. You can hire a team, customize each employee's role and instructions, assign real project tasks, and work from your own laptop in the office. Employees use AI for assigned work, conversations, and meetings; ordinary office movement makes no AI calls. Choose an installed Codex or Claude Code login, a hosted API provider, a local model, or demo mode with simulated tasks.

[!NOTE] These screenshots show Deskly running on Windows with a demo team. Demo mode lets you explore the office without an AI connection; connect a supported provider to give employees real project work. See release readiness for verified checks and platform availability.

Inside Deskly

_{Call the whole team to the boardroom, take your seat, and lead the meeting.}

_{One conversation for the room, with multiple employee replies visible together. Choose who replies and save shared decisions. This capture uses actual demo-mode replies; selected speakers reply in sequence.}

See the redesigned menu and team settings

Explore the current office and updated employees

Find rooms and keep working after dark

What is working today

Each role ships with a detailed default playbook, and you can replace it for any employee in Team & AI keys → Work instructions. Read the agent architecture for memory, coordination, and cost details.

The first-day introduction includes Try a guided project task: check one connection, assign a small task, review permissions and inspect its result. For help, see the user guide. Settings → Privacy & data can save a diagnostic report with versions and counts, excluding keys, private text and paths. Privacy and storage details.

The original single-floor office is retained, with outdoor grounds, streets, parked cars, trees, courtyard seating and a fountain. Outdoor grounds and performance details. Automatic rendering scale, shared outdoor geometry, batched trees, fewer simultaneous lights and reduced distant skeleton and HUD updates help keep movement smooth. Detailed employee models, face and hair customization, corrected walking and neck poses, staggered recalls, glass skylights, refined CEO furniture, CLI sign-ins, resumable tasks, verified memory, structured project areas and office sounds remain available. The current source passes 111 automated checks. Desktop checks cover the guided first task and a discussion with all 15 employees, selected replies, shared decisions, a two-person CEO discussion, a missing API key, and reopening history without renderer errors.

Lead a shared discussion

Press C or M, choose a room and participants, and call them over. Press C or M again to open the shared conversation. Post to everyone saves your message without calling AI. Select speakers and choose Discuss with selected for one reply each, in order; later speakers see earlier replies. Leave the composer empty to discuss your latest posted message. Nothing starts an endless automatic conversation.

Use Save shared decision for facts the group has agreed on; these enter each participant's verified memory for the selected project. Ordinary suggestions stay in the conversation history. Stop discussion cancels the round; End and save meeting keeps the transcript and existing room notes. The saved conversation selector shows discussions with the same room and participants in the current project. This feature discusses work; assign executable tasks through the employee task controls.

Get started

Install on Windows

Download Deskly 0.1.0 for Windows (64-bit), run the installer, then follow the first-day guide. Node.js is not required for the installed app. Start in demo mode or connect an AI provider for real project work.

The installed Windows app checks GitHub Releases at startup and every six hours, downloading newer versions in the background. Choose Restart to update when ready; active tasks, conversations and commands block installation. Save unsaved editor text and settings first. Settings → Updates includes a manual check and release notes.

This early-access installer is unsigned and may show a Windows publisher warning. macOS and Linux packages are not verified or published. See release notes and checksums and current limitations.

Run from source

You need Node.js 22.13+ and npm. Windows is the currently tested desktop target; macOS and Linux package targets are configured but still need verification. Linux API-key storage requires an unlocked Secret Service keyring (for example GNOME Keyring or KWallet); Deskly refuses plaintext key storage.

git clone https://github.com/Priyanshu-1622/Deskly.git
cd Deskly
npm ci
npm start

On first launch, the setup wizard asks for your name, company, project folder, team, and laptop assistant. Start in demo mode to explore without an API key. To make employees do real AI work, choose a provider in Team & AI keys. Codex CLI and Claude Code can use their installed sign-ins after you install and sign into the respective CLI; use Check installed login to verify. Other hosted providers use an API key, and Ollama uses a local endpoint. API keys are stored through Electron's encrypted OS storage when available.

When you assign the first task in a project, Deskly creates deskly.project.json in that project's root. A new empty project also gets frontend/, backend/, shared/, docs/, and operations/. For an existing project, the map uses recognized folders such as apps/web and apps/api and leaves existing files in place. You can edit the paths in deskly.project.json for your layout; the next task reads the updated map. Each task card shows its assigned area. New files must go in that area or an agreed shared/configuration path; employees can still edit existing files when integration requires it.

npm run check     # lint plus automated runtime and UI tests
npm run smoke     # isolated desktop rendering check
npm run release:check # source release checks; no packaging
npm audit         # include desktop/build dependencies
npm run dev       # app with developer tools
npm run dist:win  # build the Windows installer into dist/

The repository also declares dist:mac and dist:linux targets. Run those on their respective platforms once packaging has been verified there.

Source pushes run checks without packaging or uploading an app. Packaging workflows require a manual dispatch. Installers are distributed as GitHub Release attachments; generated packages and old unpacked apps are excluded from the source repository.

Controls

How work moves through Deskly

You assign a task
      ↓
The employee plans and checks relevant project memory
      ↓
The employee works in your project folder and shares file claims / handoffs
      ↓
Commands or external actions wait for your approval
      ↓
You review the result, touched files, and audit trail

Idle office animations make no AI calls. Assigned work, employee conversations, meetings, and the laptop assistant can use AI. Plans and short conversations can use a cheaper model, while execution uses the employee's main model. Usage metering and a monthly ceiling help control cost; actual spend still depends on the provider and tasks. Installed CLI connectors currently start a fresh process for each model turn, so long tasks can have substantial prompt overhead.

Safety and current limits

• Project file tools stay within the chosen folder and reject symbolic link paths and common secret filenames. This does not sandbox the entire app.
• API keys are encrypted with Electron safeStorage and bound to their provider and endpoint. Changing the endpoint requires a matching key; a previous key is not forwarded. Saving a key fails if secure storage is unavailable.
• Approved shell commands run with your OS permissions and can reach outside the project folder. Review the full command and working directory. Secret environment variables are excluded; cancellation and quit stop the process tree.
• Email, publishing, deployment, and payment requests are approval-gated and recorded. Deskly does not execute those external actions yet.
• Four textured, skinned character models and outfit variants are available, with shared distance detail and procedural animation. Facial animation and authored movement clips remain planned improvements.
• New offices request approval for file writes by default. Sensitive configuration and existing files outside the assigned area always require review.
• Damaged local data is preserved for recovery; interrupted checkpoints expire after three days. Privacy settings can erase all Deskly data while keeping project files.
• This first version is early access. Start with a small project, keep backups and review generated work. Current limitations and the release/update policy explain platform support, local storage, provider validation and release checks.

Deskly now includes detailed, customizable employee models with textured faces, eyes, hair and clothing; authored exterior trees; and a full roof with glass skylights. The CEO workstation has detailed furniture, dedicated leather maps, filtered reflections and sun shadows. Further wardrobe, facial animation and wider furniture refinement are in progress. See the realism direction and asset credits / character intake for the current scope and source assets.

Project map

The agent loop and its current limitations are described in AGENT_ARCHITECTURE.md. Contributions and issue reports are welcome.

Contact

For questions, support or feedback, email desklymanager1@gmail.com. You can also report a bug on GitHub. Include your operating system, Deskly version and steps to reproduce the problem; keep API keys and private project data out of reports.

License

MIT © Priyanshu Patel

Office map and brainstorming

Click the bottom-left map or press G for the office atlas: room, teammate and facility filters, sharp floor plan, drag to pan, scroll to zoom, player locator and walking routes that remain on the minimap. B opens whiteboards, N the team, T tasks, H / F1 the guide, and Backspace clears your route.

Interact with an office whiteboard using E, or choose Whiteboards on the laptop or map. Add movable, resizable notes, images and links. Boards save locally and show their contents on the physical office boards. Export a JSON backup from the editor.

To keep everyone at the office overnight, use Tab → Team → Call everyone back · keep here. This hold continues across nights and restarts until you send them home; idle workers do not make AI calls.

See the Claude audit remediation for the source fixes, verification and remaining validation gates. See current release readiness for verified source checks and remaining release gates, and the public release plan for the broader roadmap.