Saurabh Das

Innovator,Open Source Software Evangelist,Photographer,International Correspondence Chess Player

4 minute read

After spending months experimenting with lightweight local LLMs directly on a Raspberry Pi, moving to an autonomous, always-on personal assistant felt like the logical next step. A simple chat window is fine for one-off questions, but an active agent running around the clock—proactively fetching data, scheduling tasks, and managing workflows—completely changes how you interact with AI.

Here is the story of how my self-hosted setup evolved from an early deployment into a rock-solid daily driver.

Phase 1: Experimenting with OpenClaw and Telegram

I began my journey with OpenClaw, a free and open-source autonomous artificial intelligence agent originally developed by Peter Steinberger in November 2025. It is designed to execute tasks via large language models, using messaging platforms as its main interface.

Connecting OpenClaw to a Telegram bot interface backed by OpenRouter gave me the ideal testing sandbox. OpenRouter made it easy to switch models on the fly, comparing output quality, latency, and cost across different frontier and open-weight architectures.

Within days, the setup became central to my daily routine:

  • Real-Time Data Feeds: Stock market movements, tech industry roundups, and local weather forecasts.
  • Domain-Specific Calculations: Astronomy visibility windows, planetary rise/set times, and coordinate tracking equations.
  • Hobby Tracking: Live chess updates, tournament schedules, and round-start reminders.
  • Knowledge Capture: Instant note-taking, idea dumps, and quick summaries.

The Breaking Point: Context Bloat and Fragile Upgrades

While OpenClaw proved the value of a continuous assistant, operational friction quickly mounted. OpenClaw relies heavily on a local skills system for tool calling, but this eventually led to issues in my setup:

  • Hanging Threads & Execution Errors: Tool calls often stalled mid-execution, requiring manual process restarts.
  • Context Drift: Multi-turn conversations occasionally mismatched prior state, leading to fragmented or repetitive answers.
  • Inflated Token Consumption: Even for straightforward, low-context prompts, token usage on OpenRouter ran surprisingly high due to bulky system prompt overhead.
  • Upgrade Headaches: Routine updates repeatedly broke dependencies. When a minor version upgrade ultimately required a complete, ground-up reinstall on the Pi, it was time to find a cleaner alternative.

The Switch to Hermes Agent by Nous Research

Transitioning to Hermes Agent by Nous Research (released in February 2026) solved nearly every friction point from day one. Built as a self-improving agent framework, Hermes features a closed learning loop where it creates skills from experience, nudges itself to persist knowledge, and builds a deepening model of the user across sessions.

What makes it perfect for a Raspberry Pi is that the Pi only runs the agent’s control loop, its SQLite/FTS5 memory, and the messaging channel, while the heavy lifting happens at your cloud LLM provider. The installation was clean, lightweight, and operational in minutes.

While the Hermes setup wizard (hermes setup) can automatically detect an existing ~/.openclaw directory and seamlessly migrate personas, memories, skills, and platform configs, I chose to skip the migration and start completely fresh. I wanted a clean slate to fully understand the new framework, so I manually recreated my tasks for market research, astronomy calculations, and other routines from the ground up.

Just like before, I integrated Hermes Agent directly with Telegram, utilizing OpenRouter as the backend LLM gateway. Telegram serves as the primary conversational command center across my phone and desktop, handling real-time push alerts, task dispatching, and bidirectional workflows.

Not only did I effortlessly recreate every single use case I had running on OpenClaw—market updates, astronomy calculations, chess tournament alerts, news, and weather—but Hermes pushed far beyond them:

What Changed Immediately

  • Rapid Recreation of Previous Tasks: Rebuilding my previous workloads from scratch—like the stock trackers and complex astronomy visibility formulas—was intuitive and highly educational. Once set up, they ran flawlessly with zero degradation in accuracy.
  • Native Telegram Gateway: Real-time push notifications, bi-directional commands, and fluid message streaming straight into private Telegram chats without socket drops.
  • Sharp Drop in Token Overhead: Hermes structures context and tool payloads far more efficiently, cutting daily token consumption noticeably for the exact same query workload.
  • Rock-Solid Stability: Zero hanging execution threads; background tasks run reliably 24/7 without locking the conversation interface.
  • Frictionless Maintenance: Updates apply cleanly without corrupting the local environment or requiring manual database wipeouts.
  • Beyond Basics — In-House GTD & Task Automation: Hermes excels at turning quick natural language prompts into a structured Getting Things Done (GTD) workflow, building knowledge notes, and generating lightweight custom programs/scripts on demand for recurring tasks.
  • Seamless Triggers & Model Switching: Background jobs, event triggers, and on-the-fly model switching operate smoothly without losing conversational state or reasoning depth.

The Value of an Always-On Assistant

Having a proactive assistant running quietly on low-power local hardware—alerting you on Telegram about tournament matches, computing visibility windows, and managing reminders before you ask—turns AI from a novelty into genuine leverage.

In upcoming posts, I will break down the exact configuration files, my automated trigger setup, and how to build a persistent, self-healing GTD pipeline on edge hardware.

Stay tuned for the next post where we dive into the specific trigger scripts, Telegram bot bindings, and configuration setups on Raspberry Pi.

You May Also Enjoy