Image: GitHub
Why it mattersA long agent run that needs an approval no longer forces you back to the laptop it started on.
A coding agent that runs for hours stops being useful the moment it needs one approval. That is worse when you are not at the computer it is on. CodeRail tries to close that gap. It is a self-hosted web app, released this month by Wiva, Inc. It puts every one of your coding-agent sessions on a browser and a phone.
The project went public on 31 October 2026 and has 101 stars on the day of writing. It is under the Apache 2.0 licence. The code is TypeScript. The README marks the current build as "0.0.1 early".
What the app does
The server runs as a Docker container on a machine you own. On each computer you want to control, you install a small bridge. The bridge connects out to the server. There are no open ports and no VPN. The server speaks to a browser and to an installable phone web app.
From either, you open a chat with a coding agent on any of your computers. The project supports Claude Code and Codex, through their Agent Client Protocol adapters. You see the agent's plan, its tool calls, its model and mode. Background tasks are visible too. When the agent asks for an approval, you allow or deny it from the browser. You can also answer a push notification on the phone.
The web app opens a real shell on the computer. The shell outlives the browser tab. You can view, edit, upload and download any file. A git view shows pull requests. Each chat gets a browser on the host computer. Everyone in the chat sees it live, and the agent can drive it too.
Who else is in the chat
The README names three roles per computer: Owner, Operator and Viewer. Roles go down to folders and agents. Every change is written to an audit log. Each computer's agent uses that computer's own claude and codex login. CodeRail itself never holds AI credentials. Each session keeps running when every browser is closed.
Mac builds are signed and notarized. The server stores the full chat history, so the service is not end-to-end encrypted. The README has a page on the security model. It asks for private disclosure of any vulnerabilities.
A team that already runs coding agents on more than one machine gets one web app for all of them. Approvals from a phone move the slowest part of an unattended run off the device. The audit log tells a reviewer which person allowed what.
Source
- wiva-hq/CodeRail on GitHub, read 7 November 2026.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
