Simon Willison releases web-edit 0.1a0, a web file editor for agents that runs over Tailscale

On 12 October 2026, Simon Willison published web-edit 0.1a0, the first alpha of a small Python app that provides a web interface for viewing and editing files in a chosen folder. The app is designed to run over Tailscale. The agent starts a server on a remote machine, and the owner opens it from a phone over the private network and edits the file by hand. According to the author, Codex and GPT-6 Astra Extra High wrote it.

What problem does web-edit solve?

Willison was working on a project from his phone through Codex Remote. At one point he needed to edit a file containing credentials, but he did not want to pass the keys through the language model itself. Asking the agent to write a secret into a file means sending it into the context, and from there it can end up in logs, session history and traces. The author came up with a workaround: the agent launches a simple file editor, the human connects to it directly and types the value in himself. The model sees that the file has changed, but it does not see the contents that were written into it.

This is a typical problem when working with agents from a phone. There is no terminal at hand, and a full SSH client is awkward on a phone. Agents constantly need tokens and keys, and passing them through a chat with the model is not safe.

How the interface works

The screenshot in the post shows a file manager titled Workspace. At the top are Refresh, Upload files and New file buttons, and below them is a breadcrumb trail such as web-edit / web_edit / static. The table has three columns: name, size and modification date. Each file has Download and Delete links. Images up to 2 MB display directly in the interface, and selecting a text file opens the editor. A separate Upload Queue panel handles uploads: you can set a destination folder, drag and drop files or pick them with a button, and the path and name are set separately for each file.

The editor is built on CodeMirror 6. Willison considered Monaco, but it still lacks a good mobile experience, and the project was built specifically for the phone.

Key details of the release:

• version: 0.1a0, early alpha;
• server language: Python;
• access: private Tailscale network, launched by the agent;
• editor: CodeMirror 6, chosen over Monaco for mobile UX;
• file upload: through a queue with an individual path and name;
• preview: images up to 2 MB display inline;
• code authors: Codex and GPT-6 Astra Extra High.

In the example in the screenshot, app.js takes up 671.5 KB, and next to it sits THIRD_PARTY_LICENSES.txt at 32.3 KB. It looks like a built frontend with third-party libraries, primarily CodeMirror, ended up in the static folder. That is our guess; the post does not confirm it.

What to take away from this story

The tool itself is modest, but the pattern is useful for anyone who gives agents access to a working machine. If a secret needs to go into a file rather than into a conversation, it is worth setting up a separate channel for the human. A simple web editor on top of a closed network solves this without complex infrastructure.

A few things are worth checking before you use it. An alpha release does not promise a mature security model, and the post does not say whether the app has its own authentication. So it is best to expose the server only inside Tailscale, limit the root folder to the minimum necessary and not run it as a privileged user. It is also worth asking the agent to shut the process down after the edit. If you are building similar mobile workflows for Codex or another agent, it makes sense to separate two streams from the start: code and commands go through the model, while secrets and configuration go through the human.

The source code and the release are published on GitHub in the simonw/web-edit repository, tag 0.1a0.