AI-Assisted Code Review in the Theia IDE

With AI coding agents, developers are producing more code than ever, and code reviews are turning into the bottleneck. The obvious next step is to let AI do the reviews too.

Most AI review tools try to replace the human reviewer. They run as a separate step: the AI produces findings up front, posts comments, and leaves you to verify whether any of it is actually correct. For teams that are accountable for code quality, that often doesn’t really save time. You end up reviewing the reviewer.

So we wanted to try a different approach: still let the AI perform a real review, including exploring the code and producing findings, but integrate that review directly into the human’s workflow, aspect by aspect, as a conversation rather than a handoff.

That’s the idea behind the PR Reviewer agent in the Theia IDE: an AI-assisted workflow where AI and human review the changes together, iteratively. The agent handles the tedious parts of a code review (checking out branches, navigating changes, drafting findings, writing comments) while you stay in control of the outcome.

Watch the Demo

The Problem with AI Code Reviews Today

A growing number of tools run AI reviews on pull requests, often triggered automatically. They scan diffs, post comments, and flag potential issues. The result is a new burden: you now have to review the AI’s review.

The findings are often noisy, lack project context, and can’t be easily steered. You didn’t ask for them, and you can’t influence what gets flagged or how. For teams where code quality matters, this doesn’t reduce work, it just shifts it.

The deeper issue is that these tools try to remove the human from the process. But in many contexts, the human is the process. What reviewers actually need is help with the tedious parts: checking out branches, navigating changes, remembering what to look at, and writing comments on the right lines.

Integrate, Don’t Replace

The PR Reviewer agent in the Theia IDE still performs a full AI review: it explores the code, reasons about the changes, and produces findings. The difference is that this review is not a separate step you have to validate afterwards. It is interleaved with your own review, one aspect at a time.

Here’s what happens when you ask it to review a pull request:

The agent fetches the PR metadata via GitHub MCP and prepares your local workspace (checking out the right branch, remembering your previous state). It then builds a review plan: an overview of all changes together with its initial findings from exploring the code. Instead of stepping through the diff file by file, the agent structures the review semantically, grouping related changes across files into aspects such as a refactored API, a new feature, or a behavioral change. It then walks you through these aspects one by one. For each aspect it shows you its findings, you discuss them, you add your own observations, and you decide together what ends up in the review. You can confirm a finding, reject it, refine it, or write your own comment. When you’re done, the agent posts the resulting review back to GitHub and restores your workspace.

The AI’s findings are a starting point for the conversation, not a verdict you have to audit. Because the review happens iteratively, your input shapes what the agent looks at next, and the final result is genuinely shared between you and the AI.

In the demo, this plays out concretely: the AI flags a removed feature as “info”, but during the walkthrough, the reviewer realizes it’s actually a bug. Removing the plus/minus button means users can no longer enter negative numbers. That’s exactly the kind of judgment that only a human reviewer catches, and exactly why the AI review and the human review need to happen together rather than one after the other.

Built on Theia AI

The PR Reviewer is built on Theia AI, the open framework for building AI-native tools. It uses the same architecture (agents, MCP integration, sub-agents) that powers Theia Coder and other domain-specific AI workflows. Everything shown is fully open source.

The review workflow is a good example of what Theia AI enables: AI that supports structured, human-driven processes, not just AI that generates code. Code review is one instance, but the same approach applies to any domain where quality, accountability, and human judgment matter.

Try It Yourself

🧠 Download the AI-powered Theia IDE

🧰 Learn more about Theia AI

📺 Watch the full YouTube demo

Alpha Release: We Want Your Feedback!

The PR Reviewer agent is shipping as an alpha in Theia IDE 1.71. It’s ready to try, but we’re still actively shaping it based on real-world use. Your feedback at this stage has the biggest impact on where it goes next.

Give it a spin on your own pull requests and let us know what works, what doesn’t, and what’s missing:

💬 Share feedback and ask questions on GitHub

Build Your Own AI-Native Tools

Everything shown in this video is open source. The PR Reviewer is built on Theia AI, a framework for building AI-native tools for coding, diagrams, configurations, or any domain-specific workflow.

At EclipseSource, we help teams design and implement AI-native tools and IDEs. We also offer AI coding training and guided adoption to help teams build effective, repeatable AI workflows, from code generation to code review.

👉 Services for Building AI-Enhanced Tools and IDEs

👉 AI Coding Training & Guided Adoption

👉 Contact us to discuss your project

💼 Follow us: EclipseSource on LinkedIn

🎥 Subscribe to our YouTube channel: EclipseSource on YouTube