Hopkins Jesse

Posted on Jun 3

I Built a Local AI Agent That Handles My Git Workflows — Here's What I Learned After 3 Months

#productivity #ai #opensource #automation

Everyone's obsessed with cloud AI agents in 2026. Claude agents, GPT-5 workflows, Copilot Studio. I went the other direction.

I built a local AI agent using Ollama + a custom Python script that watches my git repos, reviews PRs, generates commit messages, and even writes documentation. No API calls. No monthly subscriptions. Just a Raspberry Pi 5 and 16GB of RAM.

Here's the honest truth after 90 days of daily use.

The Setup That Actually Works

I started in January 2026. My stack was simple:

Ollama running CodeQwen1.5-7B-Q4_K_M
Python watchdog monitoring 12 repos
Custom git hooks for commit message generation
A meager 4GB VRAM laptop

Total cost: $0 in API fees. My electricity bill went up about $3/month.

The core loop is boring but effective:

File change detected by watchdog
Diff extracted and tokenized locally
Model generates commit message
User reviews before committing

No magic. No agentic reasoning. Just pattern matching on steroids.

Where It Shines (And Where It Doesn't)

After tracking 847 commits over 3 months, here's my accuracy data:

Task	Success Rate	Time Saved
Commit message generation	73%	2.4 hours/week
PR description writing	58%	1.8 hours/week
Code review comments	41%	0.6 hours/week
Documentation generation	34%	1.1 hours/week

The commit messages are genuinely good. My team stopped complaining about "fixed stuff" and "update" messages within two weeks.

The code reviews? Brutal honesty: they're worthless for complex logic. The model catches missing semicolons and obvious null checks. It misses 90% of architectural issues.

The Real Problem Nobody Talks About

Here's what I learned that every "build your own AI agent" tutorial skips:

Local models hallucinate differently than cloud models. Cloud models hallucinate confidently. Local models hallucinate randomly. One day it correctly identifies a race condition. The next day it suggests adding time.sleep(1) as a fix.

I had to build a fuzzy logic layer that cross-references suggestions against known code patterns. That took 3 weekends of work. Not the 2 hours Medium articles claim.

The Killer Feature I Didn't Expect

The best part isn't the git workflows. It's the pre-commit hook that runs the model on every staged change.

import subprocess
import json
from ollama import Client

def generate_commit_message():
    diff = subprocess.run(['git', 'diff', '--cached'], capture_output=True, text=True).stdout
    client = Client(host='http://localhost:11434')
    response = client.generate(model='codeqwen:7b', prompt=f'Generate a conventional commit message for:\n{diff}')
    return response['response'].strip()

if __name__ == '__main__':
    message = generate_commit_message()
    print(f'Suggested commit message:\n{message}')
    approval = input('Accept? (y/n): ')
    if approval.lower() == 'y':
        subprocess.run(['git', 'commit', '-m', message])
    else:
        print('Manual commit required.')

This runs in 200ms on my machine. That's faster than any cloud API call including latency.

The Brutal Numbers

After 90 days:

847 commits generated
618 accepted without edits (73%)
47 false positives caught by unit tests
0 production incidents caused by AI suggestions
3 incidents where I accepted a bad commit message that confused my team

The false positives are interesting. The model suggested adding import os to a file that already had it. Another time it renamed a variable from user_id to userId in a Python file. These are harmless but annoying.

Why Nobody's Talking About This

Three reasons:

It's not scalable — This works for my 12 repos. For a monorepo with 200 developers, you need a different architecture.
It's not impressive — Cloud agents generate entire features. My thing writes commit messages. That's boring. But boring works.
There's no business model — I can't sell this. It's a shell script with a model. Nobody wants to pay for that.

The AI agent hype train in 2026 is all about autonomous coding agents that build entire apps. Meanwhile, the practical stuff that saves developers 5 hours a week is too mundane to tweet about.

The One Thing I'd Change

If I rebuilt this today, I'd add a feedback loop. Every time I reject a commit message or edit a PR description, I'd log that to a SQLite database and fine-tune the model weekly.

I haven't done it because it's another 2 weekends of work. But the data is sitting there. 229 rejected commits with my corrections. That's a goldmine of training data.

Should You Build This?

If you have a Raspberry Pi 5 or a spare laptop, yes. The setup takes 4 hours including model download. You'll save that time back in 2 weeks.

If you're expecting AGI in your terminal, no. This is a tool, not a replacement.

Here's my honest take: the most useful AI tools in 2026 won't be the ones that replace developers. They'll be

💡 Further Reading: I experiment with AI automation and open-source tools. Find more guides at Pi Stack.

💰 Want to make some smart bets? I've been using Polymarket — the world's largest prediction market platform — to bet on everything from election outcomes to tech trends. Real money, real probabilities, real payouts. Unlike crypto casinos, Polymarket is a legitimate information market where your edge comes from being better informed than the crowd. I've banked some solid wins calling AI regulation timelines and crypto ETF approvals. Sign up with my referral link and start trading: Polymarket.com

DEV Community