Email Triager

Context

My entire feed was completely dominated by OpenClaw for the last couple months and I decided to give it a try. Because at the moment, I’m very much a broke university student with a gaming PC, I decided to try my hand at running it with a local model via Ollama. Unfortunately, fate had other plans for me.

Seeing that I hadn’t made my monthly LinkedIn post to keep up my visibility, I decided to just turn it into a simple Python script and keep the streak going.

That, and I found the idea of reducing the toll on my morning mind when my anti-OCD software hadn’t actually booted yet.

Implementation

1. Gmail API

Initially, I opted to use the gogcli tool as I had some experience with it previously. However, after I was 90% through the project, I found out that it actually couldn’t modify the received emails. It could send, but it couldn’t modify. Quite an odd limitation to have.

So I ended up using the Gmail API through Google Cloud.

2. Classification

Since I wanted to use Ollama for the classification, I did end up using gemma3:1b —one of the better models my computer could run without screaming— to actually classify the email, and once classified as either important or not, it then marks it as read or keeps it as unread.

def ollama_classify(email_text, model="gemma3:1b"):
    prompt = f"""
    You are an email triage assistant. Read the following email content and classify it as important or not.

    Rules:
    - Only mark emails from real people as important.
    - Newsletters, marketing emails, or automated notifications should not be marked important.
    - Output must strictly follow this format:

    Result: True or False
    Reason: <reasoning text if True, otherwise leave empty>

    Email content:
    {email_text}
    """

    response: ChatResponse = chat(
        model=model,
        messages=[{"role": "user", "content": prompt}]
    )

    content = response.message.content.strip()

    result_match = re.search(r"Result:\s*(True|False)", content, re.IGNORECASE)
    result = result_match.group(1).lower() == "true" if result_match else False

    reason_match = re.search(r"Reason:\s*(.*)", content, re.DOTALL)
    reason = reason_match.group(1).strip() if result and reason_match else ""

    return result, reason

Conclusion

It is quite obvious that most people who use Openclaw with local models would come to the conclusion that it’s overhyped, but when I spent 50 cents from my OpenAI account and ran it with gpt-5.4-pro, its performance was quite good. However, when I used its recommended model —glm-4.7-flash— it actually couldn’t get much done and on my hardware (Nvidia RTX4070), each request took nearly 2 minutes as evident from the first figure.

Maybe at some point local models will actually be viable, but until that point, I will most likely not use it for personal purposes for the foreseeable future.