Health Companion: Autonomous Local AI Assistant for Personal Health Management with Multi-Agent Orchestration

Chatting About This Local AI Health Companion Skill

Recently, while browsing GitHub, I came across a pretty interesting project called Sofia—a local AI assistant framework written in Go. It includes a skill named health-companion, which literally means health companion. To be honest, at first I thought it was just another AI health app that requires an internet connection, but after reading the introduction—everything runs locally, supports over 20 large models of your choice, and even enables multi-agent coordination—I was pleasantly surprised. I've always wanted a reliable yet privacy-preserving health Q&A tool, because nobody wants to casually share their health issues with the outside world.

So What Exactly Does This Skill Do?

In simple terms, it's a multifunctional AI health assistant. You can ask it questions like "What medicine should I take for a cold?", "My knee hurts after exercise recently, what should I do?", or "Help me design a weekly fitness plan." Under the hood, it has more than 40 tools, such as searching medical knowledge, analyzing symptoms, calculating calories, generating dietary advice, and more. Moreover, it's not tied to a specific large model—you can freely switch in the configuration. Want to use ChatGPT? Go ahead. Want to run Llama locally? That works too. The coolest part is that it can self-improve; the more you use it, the better it understands your needs.

The Installation Process Is a Bit Tricky, But Worth It

Installing this skill isn't too complicated, but if you're new to Sofia, you might need a little patience. Here are the steps I followed:

• Install Node.js and npm: Make sure the version isn't too old, otherwise npx might have issues. I used v18 LTS, which is rock solid.
• Install Sofia CLI globally: Run npm install -g sofia-cli in the terminal. This step is necessary to easily manage skills later.
• Add the health-companion skill: Execute npx skills add github.com/grasberg/sofia --skill health-companion. This command automatically downloads the skill files. Note that it requires an internet connection, but only for the initial download.
• Configure the LLM provider: After entering the skill directory, there's a configuration file where you need to fill in the API key for your chosen large model, such as OpenAI's. If you don't want to use a paid service, you can also let Sofia use a local model, like Mistral supported by Ollama. I personally used a local model—though it's a bit slower, it's free and preserves privacy.
• Start the skill: Finally, run sofia start --skill health-companion, and you can chat with it in the terminal.

I hit a snag during my initial installation: I forgot to install Sofia first and directly ran npx, resulting in an error saying the module wasn't found. Only later did I realize that the npx skills add command is provided by Sofia, so the CLI must be installed first. Additionally, if you're using a local model, remember to install Ollama and pull the model beforehand; otherwise, you'll get a prompt saying the model is unavailable.

Hands-on Experience: It Actually Has Something

Once installed, I couldn't wait to start testing. I asked a common question: "I've been suffering from insomnia recently. Are there any natural remedies?" It responded quite comprehensively, first suggesting adjustments to the sleep environment, reducing caffeine intake, and offering several relaxation techniques like the 4-7-8 breathing method. It didn't recommend any weird folk remedies—everything was based on modern medical knowledge. I think it's because, under the hood, it can query multiple knowledge sources, cross-verify, and then respond, making it relatively reliable.

Another aspect that impressed me was multi-agent coordination. For example, I asked, "I'm 175 cm tall and weigh 80 kg. Help me design a fat-loss plan." It broke the task into several steps: first, the nutrition agent analyzed my BMI and basal metabolic rate; then, the exercise agent recommended a schedule of 3 strength sessions + 2 cardio sessions per week; finally, the lifestyle agent reminded me to ensure 7 hours of sleep. The entire conversation felt like three experts were helping me simultaneously—pretty advanced.

Of course, there were also failures. Once I asked, "Can I drink alcohol after taking cephalosporin?" It replied, "No, it can cause a disulfiram-like reaction"—that was correct. But when I followed up with, "What if I've already consumed it?" it became a bit incoherent, saying "Please contact a doctor," and then gave a vague first-aid suggestion. Apparently, for emergency situations, AI still can't replace a real doctor. It even pointed this out itself, stating that it's just an auxiliary tool and cannot diagnose or treat. I appreciated that—it didn't just make up answers to appear smart.

A Few Tips and Precautions

• Make good use of the tool chain: Health-companion includes over 40 built-in tools. Among them, the "symptom query" tool is quite practical—enter symptoms and it returns a list of possible diseases (for reference only). But never use it as a diagnostic basis.
• Try switching models: Different models have different response styles for health questions. For example, GPT-4 is more cautious and emphasizes "consult a doctor," while a local model like Mistral might be more straightforward. You can choose according to your needs.
• Data persistence: The Sofia framework supports memory functionality; health-companion remembers your conversation history, so the next time you ask a similar question, it can refer to previous information. However, if you find the records sensitive, you can disable this feature.
• Don't trust it completely: As powerful as this skill is, it's still an AI and cannot be foolproof. Especially for medication advice, always follow a doctor's guidance. I generally use it only as a knowledge base and source of inspiration, not as a medical decision-making tool.

Conclusion (Though Not Required, Just a Few Words)

Overall, health-companion is a promising local AI health skill. Its biggest advantages are privacy protection and flexible configuration—you can run it entirely on an offline computer. Moreover, because it's an open-source project, more developers will likely contribute tools to it in the future. If you enjoy tinkering and have an interest in health management...