Co-Review

Co-Review is a code review tool that lets you read an entire repository the way you would read a pull request, with the coding agent that built the code sitting in the thread. Instead of scanning a diff and guessing at intent, you open the whole repository as a review, then comment on a single line, a function, a file or an entire folder and ask why. The agent answers from the code itself, not from its memory of an earlier chat. It is built for people who ship code written with AI agents and want to keep a real understanding of what is in their codebase. Beyond the whole-repository view, Co-Review also reviews diffs, pull requests and designs. The product comes directly from a problem its maker ran into while building with a coding agent most days. Each individual diff looked fine. Six months in, the codebase worked, but he could not tell where things lived or why they had been built that way. The code had stopped feeling like his. That gap, code that functions but is no longer understood by the person who owns it, is exactly what Co-Review targets. Reviewing a pull request the way you would normally read one gives you a line-by-line pass over a small change, but it never gives you a map of the whole system. Co-Review extends that familiar review ritual to the entire repository so the understanding stays with the person responsible for the code. The first capability is the whole-repository audit. In this mode the agent writes a guide to the repository covering the map, the vocabulary, the flows and the weak spots, with findings grouped by area. Rather than dumping a single wall of commentary, the audit organizes what it finds so you can work through the codebase area by area. Coverage tracks what you have actually read, which means the audit is not a one-shot report but a record of your own reading progress through the repository. For code that was largely generated by an agent, this turns an opaque codebase into something you can walk through systematically and finish with a real sense of how the pieces fit together. The second capability treats your own change as a PR page. Line comments, suggested edits and a single verdict come together in one place, and the agent acts on that verdict. In practice this means you review a change the way a reviewer would review a pull request, but the author on the other side is an agent that can respond and make the edits. A suggested edit is not just a note left for later; it is something the agent can carry out. The single verdict keeps the interaction decisive: rather than a long back-and-forth, you state the outcome and the agent responds to it. The third capability covers someone else's pull request. Here the agent takes the first pass, and you decide what goes back to the author. Once you have made that decision, Co-Review posts to GitHub as a real review with line comments. This keeps the human in the position of judgement while the agent handles the initial sweep of the code, and it means the output of the review lands where the author already works, as a genuine GitHub review rather than a separate document that has to be copied across. The fourth capability moves review earlier, to design before code. Co-Review produces an L1/L2/L3 plan with Mermaid diagrams, and that plan is revised from your comments before anything is built. Because the plan is diagrammed, the structure is visible rather than buried in prose, and because it is revised from your comments, the design conversation happens while changes are still cheap. Reviewing a design first means the agent builds against a plan you have already read and shaped, instead of you discovering the structure only after the code exists. Co-Review describes itself as a review tool with IDE features, and that framing explains its underlying approach. It is built on Eclipse Theia, so navigation, search, Git and language servers are the real thing rather than simplified reimplementations, and everything unrelated to reviewing has been removed. In other words, it keeps the parts of an editor that you need in order to read code properly and drops the rest, so the surface stays focused on review. The same approach carries into agent support: Co-Review is made for Claude Code with a one-plugin install and ships native packages for Pi and Oh My Pi, while any MCP agent works, including Codex, Cursor, Gemini CLI, Zed, OpenCode and Goose. Reviews live on your machine, there is no account and no telemetry, and your code goes only to the agent you chose. The benefits follow from that design. You keep authorship of agent-built code by reading it rather than merely accepting it, and you can ask why something exists at the exact place where the question arises, whether that is a line, a function, a file or a folder. The answers come from the code, so they stay accurate as the code changes rather than reflecting a stale conversation. Coverage tracking turns repository reading into something with a finish line. Reviewing your own changes, other people's PRs and designs all happen in the same tool, so the habit of review extends across the whole development cycle instead of being confined to merges. Concrete uses follow the same pattern. A developer who has spent months building with an agent can run a whole-repository audit to relearn the map, vocabulary, flows and weak spots of their own project. A developer finishing a change can open it as a PR page, leave line comments and suggested edits, and give one verdict for the agent to act on. A maintainer facing an incoming pull request can let the agent take the first pass, decide what goes back to the author, and have it posted to GitHub as a real review with line comments. A team can review an L1/L2/L3 plan with Mermaid diagrams before any code is written. And when a question comes up mid-read, commenting on the specific line, function, file or folder gets an answer grounded in the code. Co-Review is aimed at developers who build with coding agents and want to stay the author of what they ship, including users of Claude Code, Pi, Oh My Pi and any MCP-compatible agent such as Codex, Cursor, Gemini CLI, Zed, OpenCode or Goose. It is free, MIT licensed, requires no account and collects no telemetry. Installation is a single command run through the published install script, and reviews stay on your machine. The project is open source and its repository is publicly available. The takeaway is straightforward: agent-built code can work and still stop feeling like yours, and Co-Review fixes that by making the whole repository reviewable with the agent that built it. Comment on a line, a function, a file or a folder, ask why, and get an answer from the code, while diffs, PRs and designs get the same treatment.