AsyncReview is an open-source tool that performs agentic code review on GitHub pull requests and issues. Its self-description is that it uses Recursive Language Models to go beyond simple diff analysis, autonomously exploring the repository, fetching relevant context, and verifying its findings in a secure sandbox before answering. Its README states it is inspired by DevinReview.
The loop it documents has five stages: the agent reasons and plans, generates Python code, executes that code in a Python REPL sandbox alongside model queries and tool commands, has its tool calls — the README names fetching a file and searching — intercepted and served from the GitHub API, then observes the result and repeats recursively.
The argument the README makes for this shape is a contrast with tools that read only the changed lines. It presents four pairings: limited context against reading any file in the repository to understand dependencies, static analysis that guesses how code works against executing search queries and verification scripts, inventing library methods against citing existing file paths and lines, and one-shot generation against iterating before answering. This is the project’s own characterization of the alternatives, not an independent comparison. The README reports no evaluation or benchmark; the one performance claim it makes is unquantified, its architecture diagram terminating in a “10x High Quality Answer”.
Capabilities
The tool runs without installation through npx asyncreview review, taking a pull request or issue
URL and a free-text question. It requires a Gemini API key, supplied as GEMINI_API_KEY; for
private repositories it additionally needs a GitHub token, either as GITHUB_TOKEN or through a
--github-token flag. The README notes the token can be obtained from the GitHub CLI.
Beyond direct invocation, AsyncReview is designed to be used as a skill by other agentic providers —
the README names Claude, Cursor, OpenCode, Gemini and Codex — on the stated grounds that this lets
them see and reason about codebases they have no local access to. Installation for that path is
npx skills add AsyncFuncAI/AsyncReview, which the README says works with agents compatible with
vercel/skills; manual setup points the agent at a skills/asyncreview/SKILL.md file. A backend
server and web interface can also be run locally, documented separately in the repository.
Adoption & Ecosystem
The repository is published under the MIT license by the GitHub account AsyncFuncAI. Its file
listing reflects the several entry points described above, with separate cli, npx, web and
skills/asyncreview directories.
What the project represents in the wider picture is the skill-packaged form of Agentic Code Review — a reviewer that acts in the repository rather than reading a diff, offered not only as a tool of its own but as a capability another coding agent installs and calls.