pocket-desk-agent 1.0.2

Pocket Desk Agent — a secure Telegram bot that gives you full remote control of your Windows PC, powered by Google Gemini AI.

Pocket Desk Agent

Your PC in your pocket — remote control, AI automation, and developer tools — all through Telegram.

Commands • Development • Contributing • Security

English • 中文 • Русский • Español • Deutsch • Français • 日本語 • Português • 한국어 • Türkçe • Українська

Pocket Desk Agent is a self-hosted Telegram bot that gives you full remote control of your Windows PC from any device. It runs entirely on your machine — no cloud relay, no subscription, no data leaving your network beyond Telegram's message relay and the optional Gemini API.

Out of the box, with zero AI setup:

• Browse and read files sandboxed to your approved directories
• Control your desktop — screenshots, keyboard shortcuts, clipboard, window switching, sleep, shutdown
• Automate UI with OCR-based text clicking (Tesseract) and element detection (OpenCV)
• Drive Claude Desktop and VS Code remotely without touching your keyboard
• Record macros and replay multi-step workflows with a single command
• Schedule tasks to run while you sleep — survives restarts
• Build and deliver Android APKs from React Native projects via Telegram

Add Google Gemini 2.0 Flash credentials to unlock:

• Conversational AI chat with multi-turn memory and image analysis
• Agentic computer use — Gemini can browse files, take screenshots, click, type, and automate your PC from natural language, with human-in-the-loop confirmation for any destructive action
• Prompt enhancement via /enhance

---

Key Features

Everything below works with no AI configuration required:

• File System Explorer: Browse, read, and search local PC directories from your phone, sandboxed to approved paths.
• Desktop Control: Take screenshots, send keyboard shortcuts, manage the clipboard, switch open windows, blank the display with privacy mode, check battery, and trigger sleep/shutdown.
• Vision & UI Automation: OCR-based clicking via Tesseract — find and click any visible text on screen. Computer vision (OpenCV) for icon and UI element detection.
• Macro Recording: Record multi-step UI sequences and replay them with a single command.
• Claude Desktop Integration: Remote control of Claude Desktop App — send prompts, switch models, manage workspaces, and automate chat flows without touching your PC.
• VS Code / Antigravity Integration: Open folders, switch AI models, and drive the Antigravity VS Code extension remotely.
• Task Scheduler: Schedule one-shot or repeating automation flows, Claude prompts, and temporary Claude/Antigravity permission watchers. Tasks survive restarts.
• Build Automation: Trigger React Native Android builds and retrieve APKs through Telegram or large-file upload links when needed.
• Auto-Update: The bot can check for and apply updates on demand.
• Lightweight: ~55-70 MB idle RAM, <0.5% idle CPU. Heavy dependencies (OpenCV, NumPy, Dropbox) load on-demand only when their commands are used.

Optional — requires Google Gemini credentials:

• AI Chat & Computer Use: Google Gemini 2.0 Flash with multi-turn conversation, image analysis, and full agentic tool-calling. Send any text or photo directly to chat. Gemini acts as an autonomous agent that can natively browse your files, analyze screenshots, and use UI automation (click, type, navigate) to perform tasks on your PC in response to natural language requests. All destructive or system-altering actions require explicit human-in-the-loop confirmation via Telegram buttons.
• Prompt Enhancement: Use /enhance to let Gemini rewrite and improve a prompt before sending it anywhere.

---

How It Works

Pocket Desk Agent runs as a local process on your Windows PC and connects outbound to Telegram's servers via long-polling — no inbound port forwarding, router configuration, or dynamic DNS is required.

Your Phone → Telegram servers → (outbound polling) → Pocket Desk Agent (local) → PC action → Reply

When you send a message from your phone, Telegram holds it until the bot's polling loop picks it up (typically under 1 second). The command runs locally on your PC and the result is sent back through the same Telegram relay.

Key internal components:

Component	Role
python-telegram-bot	Async Telegram client — receives and dispatches all commands
GeminiClient	Manages Gemini API sessions, multi-turn history, and tool-calling
FileManager	Sandboxed file I/O — all paths are validated against APPROVED_DIRECTORIES
AuthManager	Multi-provider OAuth wrapper for Antigravity, Gemini CLI, and API key modes
SchedulerRegistry	Persists scheduled tasks to disk and checks every 5 s; survives restarts
RateLimiter	Per-user token-bucket rate limiter applied automatically to every command

All 73 command handlers are registered centrally in command_map.py. Every handler is wrapped by @safe_command, which enforces authorization, rate limiting, and error reporting in a single place — no manual auth checks are needed in individual handlers.

---

Platform Compatibility

Feature	Windows	macOS / Linux
File system (browse, read, search)	✅	✅
AI chat & image analysis (Gemini)	✅	✅
Task scheduling	✅	✅
Auto-update	✅	✅
Screenshots	✅	✅
Keyboard shortcuts (/hotkey)	✅	⚠️ partial
Clipboard read/write	✅	⚠️ partial
Battery status	✅	✅
UI automation (OCR click, find text)	✅	❌
Element detection (OpenCV)	✅	❌
Window management (/windows, /focuswindow)	✅	❌
Claude Desktop integration	✅	❌
VS Code / Antigravity integration	✅	❌
React Native build automation	✅	❌
Automatic startup after login	✅	❌

macOS/Linux users can run the bot for file system access, Gemini AI chat, and scheduling. UI automation features require Windows and, for OCR commands, Tesseract.

---

Before You Start

You only need two things to get started. Google credentials are optional and only needed if you want AI chat.

1. Create a Telegram Bot

1. Open Telegram and message @BotFather
2. Send /newbot and follow the prompts to name your bot
3. Copy the bot token (looks like 123456789:ABCdef...) — this is your TELEGRAM_BOT_TOKEN

2. Get Your Telegram User ID

1. Message @userinfobot on Telegram
2. It will reply with your numeric user ID — this is your AUTHORIZED_USER_IDS

Only Telegram accounts listed in AUTHORIZED_USER_IDS can control the bot. Keep this to yourself.

3. (Optional) Get Google / Gemini Credentials

Only needed if you want AI chat, image analysis, or the /enhance command. All other features work without this.

Choose one option:

Provider note: Gemini CLI OAuth needs no project setup. Antigravity OAuth uses Google's internal code-assist API, usually auto-fetches a project, and supports GOOGLE_PROJECT_ID override if needed.

Option A — OAuth (Recommended, zero config): The bot includes built-in OAuth support — no separate GCP project or API key required for the recommended browser-login flow. During setup, choose Antigravity OAuth or Gemini CLI OAuth, or choose Setup Later and authenticate anytime via /login in Telegram.

Option B — API Key: No login flow, just paste a key.

1. Go to Google AI Studio
2. Create an API key — this is your GOOGLE_API_KEY

Custom OAuth app: If you want to use your own GCP OAuth credentials instead of the built-in ones, set GOOGLE_OAUTH_CLIENT_ID and GOOGLE_OAUTH_CLIENT_SECRET in your config. The redirect URI to register is http://localhost:51121/oauth-callback.

---

Quick Start & Installation

System Requirements

• Python 3.11+
• Windows 10 or later — required for UI automation features (pywinauto, pyautogui, pygetwindow). File system access, Gemini AI chat, and task scheduling also work on macOS/Linux (see Platform Compatibility).
• Tesseract OCR — needed for /findtext, /smartclick, and Claude/Antigravity UI automation. pdagent detects if it's missing on first run and offers to install it automatically via winget. Run pdagent setup at any time to re-check.
• Visual C++ Redistributables — required by pywinauto and pyautogui on Windows. Usually already present; install the latest from Microsoft if you hit ImportError on startup.

Option A: Install from PyPI (Recommended)

pip install pocket-desk-agent
pdagent

On the first run, pdagent launches an interactive setup wizard that walks you through all configuration values and offers to install Tesseract OCR automatically. That's it.

pdagent start        # run as a background daemon instead
pdagent configure    # re-run the setup wizard at any time
pdagent setup        # re-check and install system dependencies (e.g. Tesseract)
pdagent startup status
pdagent startup configure

Option B: Local Developer Mode

git clone https://github.com/techgniouss/pocket-desk-agent.git
cd pocket-desk-agent
pip install -e ".[dev]"
pdagent

For the full local development guide (virtual environment setup, live reloader, make targets, resource profile), see docs/LOCAL_DEVELOPMENT.md.

---

Configuration

Using the Setup Wizard (Recommended)

pdagent configure

This walks you through setting all required values and saves them to ~/.pdagent/config.

On Windows, the wizard also offers optional automatic background startup after login. This is disabled by default. If you enable it, Pocket Desk Agent starts automatically in the background after you sign in, so you do not need to run pdagent start after every reboot. This is not a Windows Service and is designed to preserve screenshot, OCR, Claude Desktop, and VS Code automation features.

The Gemini authentication step offers four choices:

Choice	Description	Best for
1) Antigravity OAuth	Opens browser immediately for sign-in using built-in credentials. Token auto-refreshes.	Most users — zero config, longest session lifetime
2) Gemini CLI OAuth	Browser login against the public Gemini API. No GCP project required. Token auto-refreshes.	Users already on the Gemini CLI ecosystem
3) API Key	Paste a Google AI Studio key. No login flow or browser needed.	Automation, headless servers, or API key preference
4) Setup Later	Skips Gemini setup. Start the bot and authenticate via /login in Telegram at any time.	Trying the bot without AI first

Token refresh: OAuth tokens (options 1 and 2) are stored locally in ~/.config/ and refresh automatically in the background. You will only need to re-authenticate if you explicitly log out or if the refresh token expires (typically after 7 days of inactivity).

Switching providers: You can switch between auth modes at any time. Run pdagent auth and choose "Switch Provider", or use /logout then /login in Telegram to start a fresh auth flow.

Manual Configuration

Copy the template and edit it:

cp .env.example .env

Required variables:

TELEGRAM_BOT_TOKEN="your_telegram_bot_token"
TELEGRAM_BOT_USERNAME="your_bot_username"
AUTHORIZED_USER_IDS="123456789"
APPROVED_DIRECTORIES="C:\Users\YourName\Documents"

Google/Gemini authentication — optional, required only for AI chat and /enhance:

# Option 1: OAuth (recommended — built-in credentials, no GCP setup needed)
# Use /login in Telegram to authenticate after starting the bot
GOOGLE_OAUTH_ENABLED=true

# Set GEMINI_AUTH_MODE=antigravity for the internal code-assist API
# or GEMINI_AUTH_MODE=gemini-cli for the public Gemini API.
# If Antigravity project auto-detection fails, set GOOGLE_PROJECT_ID.

# Option 2: Direct API key
# GEMINI_AUTH_MODE=apikey
GOOGLE_OAUTH_ENABLED=false
GOOGLE_API_KEY="your_google_ai_studio_key"

# Optional override: bring your own GCP OAuth client
# Register redirect URI: http://localhost:51121/oauth-callback
GOOGLE_OAUTH_CLIENT_ID="your_client_id"
GOOGLE_OAUTH_CLIENT_SECRET="your_client_secret"

Optional variables:

Variable	Default	Purpose
GEMINI_MODEL	gemini-2.0-flash	Gemini model to use
GEMINI_AUTH_MODE	antigravity	Auth provider: antigravity, gemini-cli, or apikey
GOOGLE_PROJECT_ID	(unset)	Optional Antigravity project override if auto-detection fails
UPLOAD_EXPIRY_TIME	1h	TempFile upload expiry for large-file links (1h/12h/24h/72h)
LOG_LEVEL	INFO	Logging verbosity
MAX_TOKENS_PER_REQUEST	8000	Gemini token limit per request
CLAUDE_DEFAULT_REPO_PATH	~/Documents	Default repo root for Claude CLI integration
SYSTEM_PROMPT	—	Custom Gemini system prompt
DROPBOX_ACCESS_TOKEN	—	Dropbox token for optional large-file uploads (see docs/dropbox-setup.md)

Legacy Config Aliases

If you are upgrading from an earlier version of Pocket Desk Agent, the following environment variable names still work but are deprecated. Migrate to the current names when convenient.

Legacy name	Current name
ALLOWED_USERS	AUTHORIZED_USER_IDS
SYSTEM_INSTRUCTION	SYSTEM_PROMPT
ANTIGRAVITY_ENABLED	GOOGLE_OAUTH_ENABLED
DEFAULT_REPO_PATH	CLAUDE_DEFAULT_REPO_PATH

Config File Locations (precedence order)

1. Shell environment variables (highest priority)
2. ~/.pdagent/config (INI format — written by pdagent configure)
3. ~/.pdagent/.env (legacy dotenv)
4. .env in the current working directory (legacy dotenv)

---

Running the Bot

Command	Description
pdagent	Run in foreground (attached to terminal)
pdagent start	Start as background daemon
pdagent stop	Stop the background daemon
pdagent restart	Restart the daemon
pdagent status	Check if the daemon is running
pdagent configure	Run the interactive setup wizard
pdagent setup	Check and install system dependencies (e.g. Tesseract OCR)
pdagent startup enable	Enable automatic startup after Windows login
pdagent startup disable	Disable automatic startup after Windows login
pdagent startup status	Show automatic startup status
pdagent startup configure	Interactively configure automatic startup
pdagent auth	Manage Gemini authentication credentials
pdagent version	Print the installed version

---

Commands Quick Reference

For the complete reference with all 73 built-in commands, see docs/COMMANDS.md.

Expand cheat sheet

Authentication & Core

Command	Description
/start	Initialize the bot
/help	Show the help menu
/status	Check Gemini API and session status
/login	Choose an authentication method and start the OAuth login flow
/authcode <code>	Enter an OAuth verification code
/checkauth	Check current authentication status
/logout	Sign out of Google
/new	Clear chat history and start fresh
/enhance <prompt>	Let Gemini improve a prompt
(any text/photo)	Chat with Gemini 2.0 Flash

File System

Command	Description
/pwd	Show current directory
/ls [path]	List files
/cd <path>	Change directory
/cat <file>	Read file contents
/find <pattern>	Search files by glob
/info <path>	File/folder metadata

Desktop Control

Command	Description
/screenshot	Capture the current display
/hotkey <keys>	Send a keyboard shortcut (e.g. ctrl+c)
/windows	List open application windows and let you switch by number
/focuswindow <number>	Activate a window from the most recent /windows list
/clipboard <text>	Set the clipboard
/viewclipboard	Read the clipboard
/battery	Battery status
`/privacy <on	off
/sleep	Put PC to sleep
/shutdown	Shut down the PC
/wakeup	Wake-on-LAN instructions
/stopbot	Stop the bot process

UI Automation

Requires Tesseract OCR. Run pdagent setup to install.

Command	Description
/smartclick <text>	Find text on screen and click it
/findtext <text>	Locate text on screen (returns coordinates, no click)
/clicktext <x> <y>	Click at specific coordinates
/findelements	Detect and number all visible UI icons
/clickelement <id>	Click a detected element by number
/typeenter <text>	Type text and press Enter
/pasteenter	Paste the current clipboard contents and press Enter
/scrollup / /scrolldown	Scroll the active window

/windows and /focuswindow switch top-level application windows. They do not switch browser tabs inside a single app.

Macro Recording

Command	Description
/savecommand <name>	Start recording a custom macro
/done	Finish recording and save
/cancelrecord	Discard the current recording
/listcommands	List all saved macros
/deletecommand <name>	Delete a saved macro

Claude Desktop Integration

Command	Description
/openclaude	Launch Claude Desktop
/stopclaude	Kill Claude Desktop
/clauderemote	Open cmd at default repo path and run claude remote-control
/claudeask <prompt>	Send a detailed prompt to Claude Desktop
/claudechat <prompt>	Automated Claude chat flow
/claudenew	Start a new Claude chat session
/clauderepo <path>	Sync a repository with Claude
/claudebranch	Claude branch management
/claudelatest	Get the latest Claude response
/claudesearch <query>	Search Claude chat history
/claudeselect	Select Claude workspace
/claudemode	Switch Claude mode
/claudemodel	Switch Claude model
/claudescreen	Screenshot of the Claude app
/claudeschedule <HH:MM> <text>	Schedule a Claude prompt

VS Code / Antigravity Integration

Command	Description
/openantigravity	Open VS Code with Antigravity
/antigravitychat	Focus the Antigravity chat panel
/antigravitymode	Switch Antigravity mode
/antigravitymodel	Switch the Antigravity AI model
/antigravityclaudecodeopen	Open the Claude Code panel in VS Code
/antigravityopenfolder <path>	Open a folder in VS Code
/claudecli [path]	Open Claude Code CLI in a folder
/claudeclisend <text>	Send a prompt to an active Claude CLI session
/openbrowser [browser]	Open Edge, Chrome, Firefox, or Brave

Scheduling

Command	Description
/schedule <HH:MM>	Start recording a scheduled automation sequence
/repeatschedule every <interval> for <duration>	Record an automation sequence that repeats for a limited time
/watchperm <claude|antigravity> every <interval> for <duration>	Repeatedly scan Claude or Antigravity for approval buttons and click them when safe
/claudeschedule <HH:MM> <text>	Schedule a Claude prompt
/listschedules	View all pending scheduled tasks
/cancelschedule <id>	Cancel a scheduled task

Build Automation

Command	Description
/build	Start a React Native Android build
/getapk	Download the latest built APK

---

Security

Pocket Desk Agent runs entirely on your local machine — no data is sent to any third-party server beyond Google's Gemini API and Telegram's message relay. Configure it carefully, as it provides system-level access to your workstation.

1. User allowlist: Every request is checked against AUTHORIZED_USER_IDS. Unrecognized Telegram accounts are silently rejected — no error is returned to the sender.
2. Directory sandboxing: File operations are restricted to APPROVED_DIRECTORIES using Path.relative_to() validation. Path traversal attacks (../) are blocked at the framework level.
3. Rate limiting: All commands are rate-limited per user. Sensitive or expensive operations have stricter limits than routine commands.
4. Secret isolation: Config and client credentials live in ~/.pdagent/, while OAuth tokens are stored in provider-specific config directories such as ~/.config/antigravity-chatbot/ and ~/.config/pdagent-gemini/, all with restricted file permissions. Never commit .env files, OAuth token files, or credential files.
5. AI safety: Gemini AI cannot execute shell commands directly. System automation tool access is tightly controlled, and any side-effecting UI interaction (keyboard, mouse, file modification, scheduling) triggers an inline confirmation prompt requiring explicit human-in-the-loop approval before execution.

---

Troubleshooting

Bot starts but doesn't respond to messages

• Confirm your Telegram user ID is in AUTHORIZED_USER_IDS (get it from @userinfobot)
• Check bot.log in your working directory for errors
• Verify the bot token is correct — a wrong token causes silent polling failures
• Run /status in the bot chat to verify the Gemini connection

/findtext or /smartclick returns an error

• Tesseract OCR is not installed or not on PATH
• Run pdagent setup to install it automatically, or manually: winget install UB-Mannheim.TesseractOCR
• After installing, restart your terminal before running the bot again

Gemini authentication fails

• Run pdagent auth and choose "Login" to re-authenticate, or use /login in Telegram
• For OAuth: make sure port 51121 is not blocked by a firewall or in use by another process; the bot starts a local HTTP server on that port to receive the OAuth callback
• For API key mode: check your key is valid at Google AI Studio
• If you see "invalid_grant", your refresh token has expired — log out and re-authenticate

Bot crashes on startup with ImportError

• Run pip install --upgrade pocket-desk-agent to ensure all dependencies are current
• On Windows, some packages (pywinauto, pyautogui) require Visual C++ Redistributables — install the latest from Microsoft's website

"Another bot instance is already running"

• Run pdagent stop to clear the stale process lock, then pdagent start
• If pdagent stop does not resolve it, find and kill the process manually: taskkill /F /IM python.exe (Windows) or check for a leftover .pid file in ~/.pdagent/

File operation fails with "Access denied" or "Path not allowed"

• The requested path is outside APPROVED_DIRECTORIES
• Add the path to your config: APPROVED_DIRECTORIES="C:\Users\YourName\Documents,C:\projects"
• Separate multiple directories with commas; use absolute paths

Scheduled tasks don't fire

• The bot must be running when the scheduled time arrives — tasks do not fire if the bot is stopped
• Run /listschedules to confirm the task is still pending and the time format is correct (HH:MM in 24-hour time)
• Check LOG_LEVEL=DEBUG output for scheduler errors

Bot is running but commands respond very slowly

• Gemini API latency can vary; non-AI commands should respond in under 2 seconds
• For UI automation, large screenshots slow OCR — try /screenshot first to confirm the display is being captured correctly
• Run pdagent status to check if a background command is still executing

/build or /getapk reports "no APK found"

• Ensure CLAUDE_DEFAULT_REPO_PATH points to a React Native project root containing android/
• Use /clauderepo <path> to set a different project directory for the session
• Check build logs in the Telegram chat for the exact Gradle error

---

Contributing

See CONTRIBUTING.md for development setup, coding standards, and how to add new commands.

---

License

Distributed under the MIT License. See LICENSE for details.