Hermes OS: A Personal Multi-Agent System for Coding, Health, and Operations

Tech blogger Tina Huang has shared her personal multi-agent system, "Hermes OS," which she uses every day. This system can autonomously build software, serve as a company COO, and manage personal productivity and health. It's not a polished commercial product — it's a fully custom-built, continuously evolving agent collaboration framework that provides an invaluable practical blueprint for anyone looking to build something similar.

Coder: The Programming Agent That Delivers Software End-to-End

The first core role in Hermes is Coder, an agent dedicated to building software. Tina interacts with it through Hermes Bot on Discord — for example, asking it to add a "skin switcher" feature to her Pomodoro desktop app.

She simply describes what she wants in plain language: add three new skins (cat, dog, and tomato) in the same style as the existing avocado avatar, let users switch between them easily, and leave everything else untouched. Within minutes, Hermes Bot produces a complete PRD (Product Requirements Document) outlining the goals, technical approach, and implementation details. After she confirms with "Looks good, let's build it," the system completes the build automatically.

Running npm start reveals that the Pomodoro app now has a skin switcher button, cycling through avocado, cat, dog, and tomato (though Tina jokes that the tomato looks "a bit terrifying"). This demonstrates a key concept: agents don't just offer suggestions — they can deliver working software end-to-end.

A Hybrid Architecture: Local Orchestration + Cloud Execution

The underlying implementation of Coder is quite clever. Hermes Bot itself is powered by a local model, Qwen3 (30B A3B), running on a Mac Studio — a popular open-source model with over 5.4 million downloads. But the actual coding work is done by having the local Qwen model "trigger" Claude Code.

The reason for this roundabout approach: Tina subscribes to Claude Max and wants to make full use of its token quota, but Anthropic doesn't allow subscription quotas to be consumed through third-party applications. So she has the local model handle orchestration and notifications, while Claude Code handles the actual writing and building, then passes results back to the Qwen model to notify the user. This "local orchestration + cloud execution" hybrid is a classic piece of engineering ingenuity for indie developers looking to work around API restrictions while keeping costs in check.

Qwen3 (30B A3B) belongs to Alibaba Cloud's Qwen series and uses a Mixture-of-Experts (MoE) architecture. "30B A3B" means 30 billion total parameters with only 3 billion activated per inference. This design allows it to run local inference on consumer hardware (like a Mac Studio) with low VRAM usage and latency, balancing cost and responsiveness. Claude Max is Anthropic's subscription plan for power users, offering a large prepaid token quota — but that quota can only be consumed through Claude's official interface or the Claude Code CLI, not via API calls from third-party apps. This is precisely why Tina needs a local Qwen model to "indirectly trigger" Claude Code. Claude Code is Anthropic's command-line coding agent that can read and write files, execute commands, and complete end-to-end code-building tasks directly in the terminal — essentially an AI programming agent with access to the local filesystem.

LifeBot: A Health and Productivity Coach Powered by Real Data

The second role in Hermes is LifeBot, which Tina interacts with via Telegram. It's a productivity and health life coach whose value lies not in "telling you information," but in "delivering personalized recommendations based on your data."

When Tina asks "how was my productivity today," it runs terminal commands and reports that she completed three 27-minute focus sessions. Ask "how many steps did I take yesterday," and it answers "11,502 steps — solid for a Sunday, above average." Even more advanced is pattern analysis: LifeBot identifies her strongest focus windows as noon, 5 PM, and 7 PM, while 9–11 AM is a "dead zone." Its recommendations are highly specific — she doesn't need more rest, but adding one extra focus session per day, especially filling in that morning gap, would push her daily deep work from roughly 90 minutes to over 110.

Three Data Sources Driving Personalized Recommendations

LifeBot's capabilities depend entirely on the data it has access to. It currently draws from three main sources: the Pomodoro app's focus logs (recording duration and task content, stored in an Obsidian vault); a custom-built Typewriter desktop app (managing to-do lists with timestamped records); and Apple Health data synced via iCloud (primarily step counts for now).

It's precisely this real, continuous personal data that allows LifeBot to give advice tailored to the individual's actual habits — not generic health tips. On the model side, LifeBot is currently powered by cloud-based DeepSeek models, with occasional fallback to local models on the Mac Studio. This reinforces a core insight: the ceiling of an agent's value is often determined by data quality, not model capability.

Obsidian is a knowledge management tool based on local Markdown files. Data is stored as plain text in a local folder, which agents can read directly via the filesystem without additional API authentication — making it a natural fit as a persistence layer for personal data. Apple Health exports (steps, sleep, heart rate, etc.) via iCloud are typically in XML or JSON format. Paired with third-party scripts that periodically parse and write structured files, this is a common low-friction path for integrating wearable data into AI systems. DeepSeek is a series of open-source large language models known for extremely low inference costs and performance approaching top closed-source models. Its API pricing is often an order of magnitude cheaper than OpenAI or Anthropic, making it a popular choice for high-frequency, cost-sensitive personal agents.

Taco Bot: The Operations Agent Serving as Startup COO

The third role, Taco Bot (modeled after takoyaki), serves Tina's company Lonely Octopus in a dual capacity: executive assistant and COO. In the team's Discord server, it sets reminders, tracks tasks, and can scan the dev channel while syncing with GitHub to report "what each team member has been working on over the past week."

The most impressive feature is the COO function. Tina openly admits that the COO role has always been the hardest to hire for at a startup — it requires documenting processes to make operations sustainable while keeping things moving fast. For years she couldn't find the right person, until she realized Taco Bot could do it. She simply says "help me draft an internal software release guide for the team," and Taco Bot searches across Discord, Notion, Slack, and other data sources to produce a pre-release checklist (e.g., at least one peer code review, automated tests) and a release execution protocol.

This document is invaluable for onboarding new team members — just point them to Taco Bot and they can learn how the team operates. Taco Bot connects Discord, Notion, Slack, Google Drive, YouTube, Instagram, databases, and various internal tools, keeping the team informed while handling both assistant and operational duties. It started on Gemini Flash (fast and easy to test) and is gradually migrating to local models to protect data privacy.

Gemini Flash is a lightweight variant of Google DeepMind's Gemini model family, optimized for low latency and high throughput. It outperforms the flagship version on response speed and API cost, making it ideal for frequently triggered tasks with relatively loose precision requirements (like task tracking and reminder notifications). Starting a prototype on a cloud-based Flash model before gradually migrating to local models is a classic incremental development strategy: "validate the logic first, then optimize for privacy and cost." Adding the operational complexity of local deployment before the core functionality is proven tends to create more problems than it solves.

The Key to Sustainable Operation: Scheduled Tasks and Alerts

Beyond the three main agents, Hermes has an "unglamorous but important" alerts channel. All running agents check in at different intervals (daily, weekly), reporting their operational status, any security vulnerabilities, and suggestions for optimization. This is the mechanism that keeps the entire system stable and sustainable over time.

For users who don't have a dedicated machine but still want Hermes running 24/7, Tina mentions that a VPS (Virtual Private Server) is a viable option — her team instance is deployed in the cloud so all members can collaborate on and extend it. She's also developing a more sophisticated agent that continuously crawls AI, tech, and education content, automatically evaluates it, and offers content ideas and software development recommendations — a truly autonomous research system.

A VPS (Virtual Private Server) is an isolated virtual machine instance carved out of a physical server by a cloud provider. Users get an independent OS environment and a fixed IP, capable of running background processes around the clock. For agent systems that need to run scheduled tasks (cron jobs) or listen for webhooks continuously, a VPS is a more stable hosting option than a personal computer. Monthly costs typically range from a few dollars to a few dozen dollars — the most common low-cost solution for indie developers to achieve "24/7 uptime." Scheduled tasks on Linux systems are typically implemented with cron or systemd timer, which can trigger scripts at intervals of minutes, hours, days, etc. Hermes's check-in and alert mechanism relies on exactly this kind of setup to have each agent report its status periodically.

Lessons for Developers and Founders

The most valuable aspect of Hermes isn't any single impressive feature — it's what it demonstrates about how an individual can build an "agent operating system" from existing tools: local models handle orchestration and privacy-sensitive tasks; powerful cloud models handle the heavy lifting; data (Obsidian, Apple Health, Notion) becomes the fuel powering agent intelligence; and Discord/Telegram serve as a unified human-agent interaction layer.

For developers and founders, this architecture proves that multi-agent collaboration is no longer a lab concept. It's a productivity tool that can be deployed daily to meaningfully improve efficiency for individuals and small teams. As open-source model capabilities continue to improve and local deployment costs continue to fall, personal AI systems like Hermes may well become standard equipment for a growing number of tech practitioners.