A new project hitting GitHub this week promises to bridge the gap between what LLMs can do in theory and what they actually accomplish on your device. RikkaHub Agent, a fork of the RikkaHub Android chat client developed by ExTV, transforms vanilla LLM assistants into genuine on-device agents with over 80 native tools at their disposal. The project targets users who want AI that actually interacts with their hardware—not just generates text responses.
Device Control That Goes Beyond Chat
The core differentiator here is real agency. While standard chat interfaces answer questions, RikkaHub Agent opens apps, sends messages, watches notifications, runs scheduled jobs, and SSHs into your servers—all via natural language commands. Users can tell the assistant to tap, swipe, scroll, type, take screenshots, adjust brightness or volume, post notifications, or read battery, WiFi, signal strength, location, sensors, contacts, and SMS data. It handles NFC tag reading and writing, Android Keystore encryption operations, external storage access, and archive management. The notification watcher feature lets you whitelist specific apps for the AI to summarize and forward incoming alerts—nothing leaves your phone until you explicitly opt in.
Telegram Integration Extends Control Everywhere
A private Telegram bot setup takes remote control to another level. Once configured, users chat with their assistant like a regular contact from anywhere in the world. Send questions, photos, PDFs, or voice notes—the latter get transcribed locally on-device using Termux integration, no cloud API required. The bot runs approval prompts through simple Yes/No buttons in Telegram chat, letting you greenlight actions while commuting, working, or sleeping. External automation tools like Tasker can also hand the agent tasks through an Intent API, slotting RikkaHub Agent into existing workflows.
Tasker-Style Automation With AI Writing Its Own Rules
The workflow system brings Tasker-style triggers and conditions under natural language control—but with a twist: the AI writes its own automation rules. Describe your desired outcome in plain English—"when I get home, turn the ringer off" or "every weekday at 8am if battery is over 50%, check my email and ping me if anything's urgent"—and the assistant generates the underlying logic. Nineteen trigger types cover WiFi, Bluetooth, headphones, geofence, app launch, notifications received, time-based scheduling, charging state, and screen on/off events. Fourteen condition types include battery thresholds, sunrise/sunset calculations, day-of-week filters, foreground app detection, and screen state checks. Receivers register only when workflows actually need them, keeping background battery drain minimal.
In-App Browser and Sub-Agents Enable Complex Tasks
The embedded browser lets the AI navigate web pages autonomously—clicking through cookie banners, filling search boxes, scrolling, and reading content back to you—all streamed via screenshots after each action. Built-in article extraction keeps token costs manageable during long browsing sessions. For computationally heavy tasks, sub-agent dispatching spawns focused side-contexts running in parallel on potentially smaller, cheaper models. One sub-agent can research flights while another updates your server simultaneously, with results consolidated into single summaries that don't clutter the main chat thread.
Three-Layer Safety Design Prioritizes User Control
Privacy and safety architecture follows a graduated approach: per-assistant toggles keep every tool disabled by default until explicitly enabled, per-call approval gates any state-changing action with options for one-time allow, session-long allow, always allow, or deny. A HARDLINE protection layer unconditionally blocks genuinely dangerous commands—wipe operations, reboots, fork bombs, system file destruction—even if users accidentally request them. Passwords and API keys never enter log files, the Telegram bot enforces an allowlist for incoming messages, cloud backups exclude server credentials and bot tokens, and notification listening starts with an empty whitelist that grows only as users explicitly select apps to forward.
The Bottom Line
RikkaHub Agent represents a significant step toward practical on-device AI agency—the kind that actually manipulates your environment rather than just describing it. The opt-in philosophy and multi-layered safety design suggest developers who understand that power without guardrails is a liability. Worth watching for anyone building mobile automation flows or exploring what LLMs can accomplish beyond text generation.