Kanjin for Codex
Codex, from a ticket to a pull request.
Kanjin runs Codex on your Mac, with the sign-in that Codex already has, such as your ChatGPT plan. It brings the ticket, a worktree in each repository, its own tools, and a diff that folds the mechanical edits. Codex writes the code, and you review it.
A window on Codex’s own app server
Kanjin starts codex app-server on your Mac, on 127.0.0.1 with a random token. The Stream view needs no terminal: it shows the messages, the approval requests and every change since the session started. The Terminal view opens Codex’s own interface on the same thread.
Join a Session continues a Codex thread of the app server, also one that you started outside Kanjin. At the end of each turn, the classifier folds the blocks that preserve behavior.
Jira tickets in Codex, without editing config.toml
Open a Jira ticket and ask Codex about it. The first message holds a brief: the description and the comments, the repositories and pull requests that name the ticket, and the pages that it links to.
Kanjin adds its own MCP server to the thread’s configuration, with a new token for each session. Your own MCP servers still load. Codex reads the ticket and its pages, and drafts a comment or a status change for your approval.
A sandbox that ends at the worktrees
Until the ticket has worktrees, Codex runs in its read-only sandbox. Implement and Fix run in a workspace-write sandbox whose writable folders are the ticket’s worktrees, and the commands get no network.
Prepare Worktrees creates the branch feature/<KEY> and a worktree in each repository that the ticket needs, and Kanjin runs each repository’s setup. Your main checkout stays as it is.
Codex code review, on any ticket
Review reads the changes of every worktree and lists defects, and Overengineer looks for needless layers. Both can run on Codex while the ticket session runs on another agent, such as Claude Code.
Fix checks each finding against the code and the ticket, fixes the valid ones and skips the rest. It does not commit. Then you read the diff, with the mechanical edits folded.
Your plan, your tools
Kanjin builds no agent of its own. It runs the Codex that you installed, so a ChatGPT plan works as it does in your terminal.
In a ticket session, Codex also gets live web search and your own MCP servers. These tools work outside the sandbox, with your accounts, and the start panel says so.
When the checks fail
Commit & Push asks Codex to commit in each worktree, and Kanjin pushes after you confirm the plan. On GitHub, Kanjin opens the pull request with your own GitHub CLI.
The ticket then shows the checks of the pull request. Fix Failures sends the failed logs and the open review threads to Codex.
Questions
Is Kanjin the Codex app?
No. The Codex app is OpenAI’s own app. Kanjin is an independent Mac app that runs the Codex CLI that you installed. Codex is a product of OpenAI.
Do I need an OpenAI API key?
No. Kanjin starts the codex command that your login shell finds, with the sign-in that Codex already has, such as your ChatGPT plan. Kanjin holds no model keys and bills no tokens.
Does Kanjin change my config.toml?
No. Kanjin adds its MCP server and the session’s limits to the configuration of each thread when the thread starts. Your configuration file stays as it is.
Which parts of Codex does Kanjin use?
The app server for the sessions, the sandbox modes for the limits of a ticket, and the thread configuration for Kanjin’s MCP server. The Terminal view is Codex’s own interface.
Can Codex review what Claude Code wrote?
Yes. Set Codex as the second agent of Review and Overengineer. It starts read-only in the ticket’s worktrees, and its report goes back to the session, where Fix applies the findings that hold.