Claude powers automated code‑review loop as GitHub project pairs AI writing with AI

Claude’s “CloCoLoop” script stitches together two of Anthropic’s and OpenAI’s flagship models—Claude Code and Codex—into a self‑correcting development pipeline, according to the open‑source repository on GitHub maintained by user amazedsaint. The workflow begins with Claude generating a feature branch, then hands the diff to Codex for critique. If Codex flags issues, Claude receives the feedback, amends the code, and the loop repeats until the review passes, at which point the tool automatically opens a pull request on GitHub. The repository includes ready‑to‑run shell scripts (`feature_loop.sh` and `review_loop.sh`) and a set of slash commands that can be installed directly into Claude Code with a one‑line curl command, making the loop accessible to developers familiar with the Claude CLI (`npm install -g @anthropic-ai/claude-code`) and the Codex CLI (`npm install -g @openai/codex`).

The implementation leans on existing tooling rather than reinventing a new AI model. By invoking Claude’s “feature‑loop” skill, developers can specify a concise task description—e.g., “Fix the session expiry bug in auth.py”—and let Claude produce the initial implementation. Codex then runs a review via the `review-loop` script, which parses uncommitted changes, generates inline comments, and returns a JSON payload that Claude consumes to produce corrective patches. The loop can be bounded by a `MAX_ITERATIONS` environment variable, preventing endless cycles, and developers can monitor progress in real time using tmux sessions that stream logs from `reviews//loop.log`. This design mirrors the pattern‑aware review approach described by Jonescodes, who built a Claude‑based DiffPrism tool that surfaces context‑aware suggestions by scanning the entire codebase rather than just the diff. While DiffPrism required a manual Claude session, CloCoLoop automates the back‑and‑forth, embedding the review directly into the GitHub workflow.

Anthropic’s broader push toward AI‑augmented development provides the backdrop for CloCoLoop’s emergence. In a recent announcement, Anthropic rolled out automated security reviews for Claude Code, targeting the surge of AI‑generated vulnerabilities (VentureBeat, 2024). The same platform now powers CloCoLoop, demonstrating how Claude’s “Model Context Protocol” (MCP) can be extended beyond single‑step prompts to orchestrate multi‑step, iterative tasks. Wired notes that Claude Code has already reshaped software engineering practices by allowing engineers to “vibe code” with minimal prompting, and CloCoLoop builds on that momentum by adding a deterministic review loop that reduces the need for human oversight in routine pull‑request cycles.

Early adopters report mixed results. The open‑source readme highlights successful runs on modest projects—adding unit tests for a payment module, fixing authentication bugs, and even large‑scale refactors—provided the codebase stays within the token limits of Claude and Codex. However, the loop’s efficacy hinges on the quality of Codex’s feedback; when Codex produces vague or overly generic comments, Claude may iterate without substantive improvement, echoing the criticism raised by Jonescodes about generic AI reviewers that miss pattern‑drift nuances. To mitigate this, the CloCoLoop scripts allow developers to run multiple loops in parallel, each isolated in its own tmux session, and to set a base branch (`BASE_BRANCH=develop`) so that the model sees the full project context rather than a shallow diff.

The practical impact of CloCoLoop remains to be quantified, but its architecture signals a shift toward fully automated code‑generation pipelines that combine generation and review in a single feedback loop. By leveraging existing Claude Code and Codex CLIs, the tool sidesteps the need for bespoke model training while still delivering a repeatable process that can be integrated into CI/CD pipelines via GitHub CLI (`brew install gh`). As Anthropic continues to expand Claude’s capabilities—most recently adding security‑focused reviews—the community‑driven CloCoLoop prototype may serve as a template for future AI‑driven development tools that aim to close the gap between code creation and quality assurance without human bottlenecks.