Open Source · Remote Compute Agent

Your Desktop,From Anywhere

A remote compute agent that gives you full AI desktop control from your phone — real-time screen streaming, text or voice input, and autonomous task execution.

Get StartedView on GitHub

How It Works

You speak, your phone understands, your desktop executes — all in one seamless flow.

Your PhoneVoice or Text
Voice Input

Open Chrome and search for flights to Tokyo

Take a screenshot and summarize what you see

Speak naturally or type — your phone understands what you want to do on your desktop.

On-Device AIUnderstands Intent

An AI model running on your phone interprets what you said and figures out if it's a simple question or a task that needs your desktop.

1Transcribes your voice to text
2Classifies: chat reply or desktop action?
3Sends command to your desktop

Simple questions are answered instantly on the phone — only desktop tasks get sent over.

Encrypted TunnelPhone → Desktop
End-to-end encrypted
Connected
Same Wi-FiFastest, direct connection
Tailscale VPNSecure mesh, any network
Cloud TunnelWorks from anywhere

Automatically picks the fastest available path. Your data never passes through third-party servers.

Safety CheckEvery Action Reviewed

Before any action runs on your desktop, it's automatically evaluated for safety.

Open VS Codesafe
Run npm testsafe
Delete system filesblocked
Run unknown scriptsandboxed

Risky commands run in an isolated sandbox — no access to your files, network, or system.

AI AgentAutonomous Execution

An AI agent on your desktop breaks down your request into steps and executes them autonomously.

Opens Chrome
Searches "flights to Tokyo"
Reads the screen to find results
Clicks on the best option
Sending progress updates to your phone
Your DesktopWindows · macOS · Linux
TerminalRun any command
GUI ControlClick, type, scroll
BrowserNavigate & interact
AccessibilityNative UI controls
File SystemRead, edit, find
SandboxIsolated environment
$ Opening Chrome...
$ Searching "flights to Tokyo"
$ Reading screen... 14 elements found▌
Your Phone

Speak naturally, AI understands

Speak or type naturally on your phone. The on-device AI figures out what you need and whether it requires your desktop.

The Bridge

Encrypted, direct, no cloud

Your commands travel directly from phone to desktop through an encrypted tunnel. Nothing is stored or routed through third-party servers.

Your Desktop

AI does the work for you

An autonomous agent takes over your desktop — opening apps, clicking buttons, running commands — while keeping you updated every step of the way.

See the animated explainer

Connectivity

Three connection paths, one seamless experience — your phone always finds the fastest route to your desktop.

LANFastest · Tried First

Same Wi-Fi Network

Direct connection, lowest latency

DiscoveryAuto-detects desktop on your network
SpeedConnects in under 1.5 seconds
PrivacyTraffic never leaves your router
Best for home & office use
TailscaleSecure · 2nd Path

Zero-Trust VPN Mesh

WireGuard-based encrypted tunnel

ReachWorks from anywhere, both devices need Tailscale
SpeedConnects in under 3 seconds
SetupAuto-detected if installed
Best for remote access
Cloudflare TunnelUniversal · 3rd Path

Public Internet Tunnel

No port forwarding, no firewall config

ReachWorks from anywhere in the world
SpeedConnects in under 5 seconds
SecurityNo open ports on your machine
Best for quick sessions anywhere
Peer-to-PeerDirect · Encrypted

Once a signaling path is found, a direct WebRTC connection is established between your phone and desktop.

1Scan QR code
2Verify with Face ID / fingerprint
3Establish encrypted channel
4Start sending commands
30sKeepalive interval
5xAuto-reconnect attempts
PairingQR Code · Biometric

Two pairing modes to fit your workflow — quick one-time sessions or persistent always-on setups.

Quick Session

Expires after 4 hours or when you disconnect

No data persisted on device

Persistent

Stays active for 30 days with auto-reconnect

Reconnects automatically

End-to-end encrypted with DTLS
Biometric verification on every pair
Old tokens auto-revoked on re-pair
CONNECTIVITY

Automatic Path Discovery

Contop tries the fastest path first and falls back automatically. If your connection drops, it reconnects with smart backoff — no manual intervention needed.

ENCRYPTION

End-to-End Encrypted

All data flows directly between your phone and desktop — encrypted with DTLS and verified with certificate fingerprints. Nothing passes through third-party servers.

AUTHENTICATION

Flexible Pairing

Quick sessions for one-time use, persistent connections for your daily setup. Each device gets one active token — re-pairing automatically revokes the old one.

Features

A powerful mobile interface designed for every workflow

Adaptive Layouts

Split ViewDefault Mode
Screen
Chat

See your desktop screen and conversation side by side. Drag the separator to resize — anywhere from 30% to 70%.

Best for monitoring tasks as they run
Video FocusWatch Mode
Full Screen
chat overlay

Maximize your desktop view. The chat floats on top as a transparent overlay — tap through it to keep watching.

Best for watching the agent work
Thread FocusRead Mode
mini screen
Full Chat

Focus on the conversation. A small video preview stays pinned at the top so you never lose sight of your desktop.

Best for reading detailed results
Side-by-SideLandscape Mode
Screen
Chat

Rotate your phone and get a widescreen view. Desktop screen on the left, conversation on the right — plus a fullscreen video option for dedicated monitoring.

Best for extended work sessions
Fullscreen VideoLandscape Mode
Full Screen

Dedicate your entire screen to watching your desktop. Minimal floating controls stay out of the way. Perfect for long-running tasks where you just need to keep an eye on things.

Best for dedicated desktop monitoring
LAYOUTS

5 Modes for Every Use Case

Split View for balanced monitoring, Video Focus for watching the agent work, Thread Focus for reading results, Side-by-Side for landscape multitasking, and Fullscreen Video for dedicated desktop viewing.

ORIENTATION

Smart Rotation

Rotate your phone and the layout adapts instantly. Set your preferred mode for portrait and landscape — Contop remembers your choices across sessions.

INTERACTION

Drag to Resize

Resize the screen and chat panels by dragging the separator. Works horizontally in portrait and vertically in landscape. Constrained so neither panel gets too small.

Intelligent Model Configuration

CONFIGURATION

Per-Role Model Selection

Three independent AI roles — conversation, execution, and screen interaction — each configurable with 25+ models from Gemini, OpenAI, Anthropic, or OpenRouter. Change from your phone anytime.

BACKENDS

Nine Screen Strategies

Nine ways for the agent to see your screen — from local OmniParser to six cloud vision models, Google's native vision, or keyboard-first with no screenshots.

RUNTIME

Switch Without Restarting

Change models and backends on the fly from mobile settings. The desktop agent picks up your new configuration on the next command — zero downtime.

Everyday Experience

LIVE THREADReal-Time Updates
Check if the API server is running
I'll check the process list...
Running command...
ps aux | grep server
Server is running on port 3000

See every step the agent takes in real time — messages, tool calls, and results stream into a live thread

Best for following the agent's reasoning step by step
SESSIONSAuto-Saved · Resumable
AllTodayThis Week
Deploy fix for API
Mar 15
12Continue
Debug login timeout
Mar 14
8
Setup CI pipeline
Mar 12
23

Pick up where you left off — sessions persist across app restarts with full conversation history

Best for resuming complex multi-step tasks
CUSTOM PROMPTYour Rules · Your Way
Always use PowerShell
My project is at C:\Dev\myapp
Respond in Spanish
Clear

Tell the agent how you want it to behave — set language, project paths, or preferred tools

Best for personalizing the agent to your workflow
DEVICE CONTROLRemote · Instant
Keep Screen Awake
Lock Desktop

Control your desktop state from anywhere

Lock your screen or keep it awake during long tasks — all from your phone

Best for hands-free desktop management
VOICE INPUTSpeak · Send · Execute
0:12
CancelSend

Speak or type — your intent becomes the command. Record voice, review, and send, or type directly for quick instructions

Best for quick commands while multitasking
MANUAL CONTROLJoystick · Hybrid Mode
L
R
EscTabDelCtrl

Take direct control — move the cursor with a joystick, click, scroll, and send key combos from your phone

Best for precision tasks the agent can't handle alone
EXECUTION

See Every Step in Real Time

Watch the agent work through your request step by step. User messages, AI responses, tool calls, and results stream into a live thread — with progress indicators and expandable details.

SESSIONS

Pick Up Where You Left Off

Every session is saved automatically with full conversation history. Browse by date, filter by tool or result, rename sessions, and continue any past session with one tap.

CONTROL

Your Desktop, Your Rules

Lock your screen, keep it awake, set custom instructions, use voice input, or take direct control with a joystick overlay for cursor, clicks, and keyboard shortcuts. Switch between AI and manual mode seamlessly.

Agent & Automation

33 built-in tools that let the agent run commands, control your screen, manage files, and automate entire workflows.

TerminalRun Any Command
$ pip install requests
✓ Installed in 1.2s

Run shell commands on your desktop just like you would in a terminal — install packages, run scripts, manage files.

SafetyDangerous commands run in a Docker sandbox
PrivacySensitive env vars like API keys are hidden
LimitsAuto-stops stalled commands after 5 seconds
Screen ControlSee & Interact

The agent sees your screen, identifies every button and element, then clicks, types, and scrolls — just like a person would.

1Takes a screenshot of your desktop
2AI identifies all clickable elements
3Performs the right action at the right spot
Click
Type
Scroll
Drag
Hotkeys
Select
Also supports keyboard-only mode via accessibility tree
BrowserNavigate & Extract

Controls Chrome directly — no screenshots needed. Navigates pages, fills forms, clicks buttons, and reads content efficiently.

→Navigate to any URL
◉Click buttons and links
✎Fill in forms and fields
≡Read page text content
◻Take page snapshots
▣Manage multiple tabs

Reads page text directly instead of taking screenshots — 10x more efficient for the AI.

FilesRead · Edit · Search

Works with any file on your machine — text, code, PDFs, images, and Excel spreadsheets.

Text & CodeRead and edit with precision
PDFsExtract content as readable text
ImagesView and analyze screenshots
ExcelRead sheets, write cells, merge ranges
SearchFind files by name or content
7 file tools available
WindowsCross-Platform

Manage windows, read the clipboard, monitor processes, and download files — works the same on every platform.

Window FocusSwitch between apps
ResizeArrange your workspace
ClipboardRead and write content
DownloadsFetch files from URLs
WindowsNative adapters
macOSNative adapters
LinuxNative adapters
Apps & SkillsLaunch · Automate · Extend

Launch and close apps, handle Save As and Open dialogs, and create reusable skills to automate repetitive workflows.

Launch any applicationWaits until ready
Close apps gracefullyAuto-saves if needed
Handle file dialogsSave As, Open, export

Custom Skills

Teach the agent new abilities by creating reusable skills — chain multiple steps into one command.

PromptWorkflowPythonMixed
EXECUTION

Three ways to control your desktop

Run terminal commands, automate GUI interactions by seeing your screen, or control Chrome directly — the agent picks the best approach for each task.

OPERATIONS

Works with any file, any platform

Read and edit code, PDFs, images, and spreadsheets. Manage windows and monitor your system. Same experience on Windows, macOS, and Linux.

EXTENSIBILITY

Teach it new tricks

Create custom skills to automate your unique workflows. Chain actions together, save them once, and reuse them forever — no coding required.

Model Providers

Use API keys or your existing subscriptions — choose from 4 providers and 20+ models, and configure any combination for any task.

ProvidersKeys or Subscriptions

Choose from 4 providers and 20+ models. Use any combination for different tasks — switch anytime from your phone.

Google GeminiFlash, Pro, and Flash Lite models for conversation, execution, screen control, and speech-to-text.
OpenAIGPT and o-series models with multimodal capabilities. Whisper for alternative speech-to-text.
AnthropicClaude Opus, Sonnet, and Haiku — with optional extended thinking for deeper reasoning.
OpenRouterUniversal gateway to 300+ models — Grok, Qwen, Mistral, Nemotron, and more via one API key.
Three AI RolesMix & Match

The app uses three independent AI roles — assign any provider to any role, and change them at runtime from mobile settings.

C
ConversationUnderstands what you want and classifies your intent
E
ExecutionRuns tools and carries out tasks on your desktop
V
VisionPicks how the AI reads your screen — multiple backends available

Any provider can fill any role — use Gemini for conversation and Claude for execution, or any other combination.

AuthenticationQR Pairing

Set up API keys or enable subscription mode on the desktop app. Configuration travels to your phone securely through QR pairing — no manual copying.

1Configure API keys or enable subscription mode in desktop settings
2Scan the QR code with your phone to pair
3Auth config is encrypted and stored in your phone's secure enclave
Gemini
OpenAI
Anthropic
OpenRouter
Models configurable per-role at runtime
CHOICE

Pick the best model for the job

Different tasks benefit from different models. Use a fast model for quick actions and a powerful one for complex reasoning — all from the same app.

CONTROL

Configure from your phone

Switch models and providers at any time from mobile settings. Each AI role can be independently assigned to any supported model.

SECURITY

Your credentials, your control

API keys and subscription preferences never leave your devices. They're configured on your desktop, transferred securely via QR, and stored encrypted on your phone.

Skills

Extensible agent capabilities via the SKILL.md standard — built-in skills included, custom skills easy to create.

Advanced Workflowsv1.0.0 · python
async def fill_form(fields) → dict
async def extract_text(region, element_name) → dict
async def copy_between_apps(source, target) → dict
async def set_env_var(name, value, scope) → dict
async def change_setting(setting_path, value) → dict
async def app_menu(app_name, menu_path) → dict
async def install_app(name, method) → dict
async def find_and_replace_in_files(path, pattern, old, new) → dict
8 Python toolspython
IDE Chatv2.0.0 · workflow
# vscode-claude-send
- action: hotkey
keys: [ctrl, shift, p]
- action: type_text
text: "Claude: {prompt}"
- action: press_key
key: enter
VS Code Claude
VS Code Copilot
Cursor
24 deterministic workflowsworkflow
Prompt SkillsSKILL.md Standard
---
name: skill-authoring
description: Guide for creating...
version: "1.0.0"
---
# Skill Instructions
Markdown body with agent...
skill-authoringGuide for creating and editing custom skills
v1.0.0
web-researchBrowser automation + Electron + CDP strategy
v1.0.0
cli-command-patternsCross-platform bash/PowerShell patterns
v1.1.0
Agent instructions loaded on demandprompt
Skill Types4 Execution Models
promptAgent instructions loaded on demand
workflowExecuted deterministically by workflow engine
pythonAsync Python functions registered as agent tools
mixedAll mechanisms available
~/.contop/skills/{skill-name}/
├── SKILL.md # YAML frontmatter + markdown
└── scripts/ # Optional
├── *.yaml
└── *.py
Skill Lifecycle5-Stage Pipeline
1
Discoverydiscover_skills() scans ~/.contop/skills/
2
Registrationenabled_skills in settings.json
3
DisclosureMetadata at startup, full on load_skill
4
ConflictsTool name check against 33 CORE_TOOL_NAMES
5
Agent Toolsexecute · load · create · edit
<!-- build_skills_prompt_section() -->
<skills>
<skill name="ide-chat" type="workflow">
Control AI coding IDE...
</skill>
</skills>
EXTENSIBILITY

Extensible Agent

Add new capabilities by dropping a SKILL.md file into the skills directory. The agent discovers and loads it automatically — no code changes needed.

AUTOMATION

Deterministic Workflows

Define YAML step sequences for repetitive tasks — keyboard shortcuts, menu navigation, form filling. Runs the same way every time, no AI guesswork.

CUSTOMIZATION

Create Your Own

Build custom skills as prompt instructions, YAML workflows, Python tools, or any combination. Manage them from the desktop GUI — discover, enable, edit.

Security

Every layer verified against the real codebase — from physical machine protection to encrypted peer-to-peer connections.

Away Modeaway_mode.rs · Physical Security
Bystander sees
Dark PIN overlay
WDA_EXCLUDEFROMCAPTURE = 0x00000011
Invisible to screen capture
Owner sees
Live WebRTC video feed
streaming
Desktop automation running
Away Mode Features3 layers
PIN-locked overlayFullscreen topmost window · low-level keyboard hook blocks everything except digit keys (0–9), numpad (0–9), backspace, and enter
Auto-engage on idleActivates after 5 minutes of no mouse or keyboard input · polls every 30 seconds via GetLastInputInfo()
3 unlock methodsScreen PIN (4–12 digits, bcrypt cost 10) · phone command (away_mode_disengage) · emergency recovery PIN (6–12 digits)
Command Classifierdual_tool_evaluator.py
Every command the agent wants to run goes through this 7-step check — top to bottom, first match wins:
1
User forced host execution→ run on host
2
Needs the screen (GUI, browser, observe)→ run on host
3
Unknown or unrecognized tool→ sandbox it
4
Forbidden command (rm -rf /, format C:, mkfs…)→ block entirely
5
Touches protected path (/root, C:\Windows…)→ sandbox it
6
Destructive (rm, kill, DROP TABLE, taskkill…)→ ask user first
7
Everything else — safe by default→ run on host
Also blocks encoded PowerShell commands · detects dangerous cmdlets like remove-item, stop-process, invoke-expression
Docker Sandboxdocker_sandbox.py
Risky commands run inside a locked-down Docker container:
No network accessContainer can't reach the internet
256 MB memory limitPrevents resource exhaustion
50% CPU capWon't slow down your machine
100 process limitNo fork bombs
Read-only filesystemCan't modify the container image
64 MB temp storageOnly /tmp is writable, and it's tiny
No privilege escalationRuns as 'nobody' with zero capabilities
Auto-starts Docker DesktopDetects if Docker is installed but not running and starts it automatically
Mobile Approvalwebrtc_peer.py
When the agent wants to do something destructive, it asks your phone for permission first:
1Desktop agent pauses
"I want to delete 3 files — approve?"
2Your phone shows a prompt
Approve or Deny with one tap
3Agent gets your answer
Proceeds only if you approved
Sent over encrypted WebRTC data channel:
{
"type": "agent_confirmation_response"
"payload": { "approved": true }
}
Audit Loggingaudit_logger.py
~/.contop/logs/session-{YYYY-MM-DD}.jsonl
timestamp: UTC ISO 8601
session_id: str
user_prompt: str
classified_command: str
tool_used: str
execution_result: str
voice_message: str (default "")
duration_ms: int (default 0)
async def log(*, session_id, user_prompt, classified_command, tool_used, execution_result, voice_message, duration_ms)
Fire-and-forget · asyncio.to_thread(self._write_line, path, line)
Auth & Encryptionpairing.py
@dataclass
PairingToken:
token
dtls_fingerprint
stun_config
created_at
expires_at
device_id
connection_type = "permanent"
TOKEN_TTL_DAYS30
TEMP_TOKEN_TTL_HOURS4
DTLS fingerprintSHA-256
STUNstun.l.google.com:19302
~/.contop/tokens.json · atomic write (.tmp → rename) · validate_token() auto-removes expired
Configurable Rulessettings.py
restricted_paths[]
/root · /etc/shadow · /etc/passwd · C:\Windows · C:\Windows\System32 · C:\Windows\SysWOW64
forbidden_commands[]
rm -rf / · mkfs · dd if= · format C: · del /f /s /q C:\
destructive_patterns[]
rm · rmdir · del · kill · taskkill · shutdown · DROP TABLE · remove-item · stop-process…
away_mode:
enabled: false
pin_hash: ""
emergency_pin_hash: ""
auto_engage_minutes: 5
idle_timeout_enabled: true
Hot-reload via mtime caching · get_settings() checks _cached_mtime · ~/.contop/settings.json
Paired Devicespairing.py · Desktop UI
See every device that can access your computer — live status, location, and connection path:
Alex's iPhoneConnected
via Local Network
Just now
Work iPadDisconnected
via Tunnel
from San Francisco, US
2 hours ago
One-click revoke — instantly disconnects and blocks the device
Alerts & Smart Pairinggeo.py · OS Notifications
Native OS notifications fire in real time — even when the app is minimized:
Device Connected
Alex's iPhone via Local Network
Token Replaced
New pairing replaced existing access
Connection path auto-classified:
Private IP (192.168.x, 10.x) → Local Network
Tailscale IP (100.64.0.0/10) → Tailscale VPN
Public IP → Tunnel — geo-located
PHYSICAL SECURITY

Away Mode

Away Mode protects your machine when you're not at the keyboard. PIN overlay, keyboard lock, idle auto-engage, encrypted secrets.

EXECUTION SAFETY

Command Classification

Every command is classified before it runs. Dangerous actions are sandboxed or blocked. You approve what matters.

CONNECTION TRUST

End-to-End Encrypted

End-to-end encrypted. Peer-to-peer. No cloud relay. Biometric pairing. Your data never leaves the tunnel.

DEVICE VISIBILITY

Paired Device Management

See every connected device, where it's connecting from, and revoke access instantly. OS alerts for every connection event.

Use Cases

Real people, real problems, solved in seconds.

Production OutageVoice → CLI
Voice Input
“Check the production logs for user-auth. If it's stalled, restart it.”
Processing
$ docker logs user-auth --tail 20
ERROR: health check timeout
$ docker restart user-auth
Up 3 seconds (healthy)
Resolved in 47s
A
Alex — Backend Engineer
On a train · PagerDuty alert firing

PagerDuty fires while Alex is on the train. He opens Contop, speaks one command, and the agent checks the logs, finds the stalled container, and restarts it. Outage resolved in under a minute — no laptop needed.

WebRTC TunnelVoice InputCLI ExecutionReal-Time Video
Security BoundaryCommand Gate
“Run render-final.sh, but first delete old temp files in root directory”
Security Gate
BLOCKEDRoot directory deletion
“Proceed with render only?”
Yes
No
S
Sarah — Motion Designer
Coffee shop · 4K render deadline

Sarah asks Contop to run her render script and casually adds “delete temp files in root.” The security gate blocks the dangerous part, asks her phone to confirm, and kicks off just the render. System stays safe.

DualToolEvaluatorSandboxUser ConfirmationRestricted Paths
GUI AutomationVisual Navigation
Blender 4.1
Render
72%
GPU Memory Error
CUDA out of memory — tile size too large
RetryCancel
Voice Command
“Lower the tile size in render settings and hit retry”
M
Marcus — 3D Artist
Dinner out · Blender render running at home

Marcus left a Blender render running on his workstation. At dinner, his phone shows a GPU memory error dialog blocking the process. He tells Contop to lower the tile size and hit retry — the agent navigates Blender's UI visually, clicks through the settings, and the render resumes.

GUI AutomationVisual StreamDesktop AppsNo CLI Needed
RESPONSE TIME

Instant Response

Resolve critical issues in seconds, not minutes. Voice-to-action from any location.

SAFETY

Safety by Design

Dangerous actions are caught, sandboxed, and confirmed before execution.

SUPPORT

Zero Walkthrough

Remote support without asking users to follow complex steps.

Download

Get Contop running on your machine in minutes.

Windows

Installer (.exe)

Download

Requires: Python 3.12+

macOS

Disk Image (.dmg)

Download

Requires: Python 3.12+

Linux

AppImage / .deb

Download

Requires: Python 3.12+

.deb also available

Or install via package manager—no security warnings

macOS:brew install slopedrop/contop/contopWindows:scoop bucket add contop https://github.com/slopedrop/scoop-contopthenscoop install contop

Android

Google Play Store

Google Play

iOS

Coming Soon

Coming Soon

Developer Documentation

Setup guides, API reference, skill authoring, and configuration.

Read the Docs
DESKTOP

Desktop Agent

Control your computer with AI from anywhere. Windows, macOS, and Linux.

MOBILE

Mobile Commander

Voice and text control from your phone. Android (iOS coming soon).

DOCS

Documentation

Setup guides, API reference, skill authoring, and configuration.