As AI coding agents proliferate across development workflows, the question of oversight has become urgent. Cotect, a newly released open-source tool from cotect.dev, aims to keep humans firmly in the loop by providing real-time visibility into what agents are actually writing to your repository.

Real-Time Code Watching

Cotect runs as a local application that monitors your working tree and Git state while coding agents do their thing. Whether you're using Claude Code, OpenAI's Codex, OpenCode, or any other agent that touches your files, Cotect captures every change the moment it happens. The interface renders file opens in approximately 20 milliseconds—a speed that's critical when agents are making dozens of modifications per session.

Hunk-by-Hunk Review Workflow

The core innovation here is hunk-level review. Instead of staring at a massive diff and trying to mentally parse what changed, developers can step through alterations one chunk at a time using WASD navigation in the terminal-based editor. Each comment you leave can be copied directly into your agent prompt of choice, creating a tight feedback loop without switching contexts.

Import Graph Visualization

Understanding codebase architecture becomes harder when agents introduce dependencies invisibly. Cotect resolves import relationships across your entire repository and renders them as an interactive graph. Click any file to see exactly what it imports and what's pulling it in as a dependency. This matters because agents write as well as you direct them, and maintaining mental model accuracy about project structure is what separates good prompts from hallucinations.

Codebase Health Metrics

Beyond individual change review, Cotect computes structural findings directly from your repository. It flags circular dependencies, identifies churn hotspots where files keep changing, and surfaces oversized files that might be straining context windows. A sortable table lets you prioritize technical debt work based on what the data actually shows rather than gut feeling.

Privacy-First Architecture

Cotect's FAQ makes a critical promise: your code never leaves your machine. The tool reads locally and renders everything on-device. The only network call it makes is to GitHub for version update checks. For organizations worried about proprietary code touching external services, this matters—a lot.

Free and Open Source

The project ships under Apache-2.0 license with no paywall on core features. The sustainable business model, per the developers, involves paid team collaboration features down the line—shared reviews, multiplayer workflows—but the essential read-only inspection tool remains free indefinitely. "If a tool reads your code all day, you should be able to read its code too," the FAQ states, pointing to github.com/cotect-dev/cotect.

Cross-Platform Availability

Cotect installs like any desktop application with launcher entries on macOS, Windows, and Linux. Portable AppImage builds are available for Linux users who prefer not to install: a single curl command downloads the binary, chmod makes it executable, and you're reviewing agent changes immediately.

Key Takeaways

  • Cotect is read-only by design—it never writes to your working tree, stage, or commits
  • ~20ms file opens enable real-time monitoring of fast-moving agents
  • Import graph visualization helps maintain architectural understanding as codebases evolve
  • Code leaves no machine—fully local processing with only version check network calls
  • Core app stays free permanently under Apache-2.0; collaboration features may become paid later

The Bottom Line

The agent coding era has arrived whether we're ready or not, and Cotect represents exactly the kind of tooling this moment demands: tools that amplify human judgment rather than replacing it. If you're delegating code generation to AI agents without rigorous review workflows, you're shipping technical debt on someone else's schedule. This thing makes oversight actually bearable.