I Built a Self-Hosted AI Engineering Team That Won't Push Code Without My Approval
Coding agents are good at writing code and bad at knowing when they're wrong. I've watched an agent confidently ship a broken change, add a dependency that was published 20 hours ago, or quietly leak a secret into a commit message. So I built AI Employee — a self-hosted system that wraps AI coding agents in the same process a careful human team already uses, before anything reaches your GitHub.…
A new self-hosted AI engineering system aims to prevent AI coding agents from making mistakes or pushing broken code without human approval. The AI Employee platform acts as a safety net before any code reaches your GitHub repository. The system breaks down the coding process into several stages, each designed to catch specific types of errors or issues.
First, a planner reads the project's code, handoff notes, and team memory to create a detailed plan for the coding task. This planning stage helps ensure the agent is focused on the right problem. Next, a quick checks phase scans the code for things like leaked secrets, broken JSON/YAML/JS files, and risky new dependencies, catching problems before they reach the coder.
The coder then creates the code in its own Git worktree, separated from the main codebase. After this, several review stages take place, including a reviewer that checks requirements, crashes, frontend/backend agreement, and tests up to three times before approving the code. UI and security critics examine the code visually and for potential security vulnerabilities, ensuring the code is safe for deployment.
Only after passing all these stages does the agent write and push a commit to your repository. However, this branch remains unmerged until a human clicks approve in the dashboard. The entire system is sandboxed, with agents running in a separate, unprivileged Linux user with no access to critical system tools or GitHub credentials. Secrets are blocked from memory and commits, and model calls are tracked for cost management.
The system is designed to be model-agnostic, allowing you to mix and match different AI models for different stages of the coding process, depending on your cost and quality preferences. The memory is stored in a single SQLite file, making backup and management simple. The developers are currently in a developer preview phase and welcome feedback and contributions to improve the review and sandboxing features.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
