Teams that cannot send source code to external services are interested in local assistants for review. They can help, if you give them the right job.
Where they help
- First-pass checklists: pointing out missing error handling, unclear names, or inconsistent style.
- Explaining unfamiliar code: summarizing what a function appears to do.
- Drafting tests: suggesting cases to consider.
- Writing commit and pull request descriptions from a diff.
Where they struggle
- Deep reasoning across many files.
- Subtle concurrency or security bugs.
- Knowing your project's conventions unless you tell them.
- Long diffs that exceed the context window.
Making it practical
- Feed small, focused diffs rather than a whole repository.
- Provide a short style guide and the most important rules in the system prompt.
- Ask for output in a fixed format: issue, location, suggested change, confidence.
- Have a human decide. Treat the output as suggestions.
Integrating with workflow
Many teams run the assistant from the editor or as an optional job in the pull-request pipeline that posts comments for the author, not as a gate that blocks merging.
Choose models carefully
Code-oriented models of medium size often offer a reasonable balance on one workstation GPU. Test on your own past pull requests where you know the real issues.
Privacy and logs
Keep the service inside your network, and decide whether prompts are logged. Code can contain secrets, so scan before sending.
For editor-based setup, see our install guide; hardware sizing is in the VRAM table; and more ideas live in the lab.
A short rollout suggestion
Start with one team and one repository for two weeks. Collect which suggestions were accepted and which were noise, then adjust the prompt and the checklist. Share a few examples of good and poor suggestions with the whole team so everyone calibrates their trust. Only then consider enabling the assistant for more projects, and keep the option to switch it off easily if it becomes a distraction.
Published by MicroAI One (本地AI 1号网), a Shenzhen team working on local LLM hardware, software and deployment.