stet

stet/stɛt/verbLet it stand.

Write papers with an agent you can

In stet, every change the agent makes is shown word by word, in the source and on the page. Nothing is in your paper until you accept it.

Runs on the plan you already haveClaudeChatGPTNo API key.

Free for Linux.

stet’s window. On the left the agent explains an edit to the abstract; in the middle the edit is shown word by word with Accept and Discard; on the right the PDF marks the changed passage.

We all use it. Might as well do it properly.

However you use AI on a paper today, the hard part is the same: knowing exactly what it changed before your name goes on it.

What you are probably doing

  1. A chat window and a clipboardIt sees one paragraph at a time, and you compare two versions by eye to find what it changed.
  2. A code editor with an agentBuilt for code. Your prose comes back as a diff of whole lines, and nothing in it knows about the PDF.
  3. A web editor with AI built inYour unpublished paper lives on someone else’s servers and runs on the model they chose.

stet

  1. The agent reads the whole paper: every section, table and reference.
  2. Every change is shown word by word, and nothing is in the paper until you accept it.
  3. The PDF rebuilds, with the changed passages marked on the pages.
  4. Your files stay in their folder, and the agent runs on your own Claude plan.
  5. Git history in the same window. Any version is one click back.
  6. References from dblp in their published version, and a check that every entry in your bibliography exists.

The agent writes. You decide what stands.

main.tex, lines 28–31

Existing systems restore the whole memory image before the first request can run, which is very slow and also wastes a lot of memory for on pages that the function never ends up touching at alltouches.

Abstract

Serverless platforms start a new sandbox whenever no warm one is available. Existing systems restore the whole memory image before the first request can run, which is very slow and also wastes a lot of memory for on pages that the function never ends up touching at alltouches. We present Kelp, which restores a function one page at a time.

Try it. This one is live.

Nothing is in your paper until you accept it

The agent edits your files, and every change waits in Review. Read it word by word, accept what is right, and discard the rest with one click.

A sentence selected on a page of the PDF.
The message box quoting the selected sentence above a question about it.

Point at a sentence instead of pasting it

Select text in the PDF or in the source and ask. The agent gets the passage and where it is, so “this” means what you selected.

History: uncommitted changes with a commit message field, above a list of earlier commits.

Every version, without the file names

stet reads the git repository your paper is in. Commit the files you choose, look at any earlier version, and restore a file from it.

Use the plan you already have

No API key and no new subscription. The agent runs on the Claude or ChatGPT plan you already pay for.

One window for the whole paper

Ask on the left, write in the middle, read on the right. Collapse a pane, swap sides, or stack them.

The agent’s account of an edit it made.

The agent

Ask for an edit or an opinion. It says what it did and why.

LaTeX source in the editor.

The source

A real LaTeX editor. Write the parts you want to write.

A page of the PDF with a changed passage marked.

The pages

Your PDF, rebuilt as you go, with what changed marked on it.

The file tree of a paper.

The files

The paper’s folder and its outline, as in Overleaf.

Why it is called stet

stet /stɛt/

verb. Latin, “let it stand”.

A proofreader’s mark. When an editor’s correction should be ignored, the author writes stet in the margin and puts a row of dots under the words to keep.

An agent is an editor that works very fast. What it changes is a proposal, and the last word is yours: accept what is right, and let your own sentence stand where it is not. That is the whole app, so that is its name.

The gap entirely closes once the image is warm.

stet

The particulars

Specifications

Version
0.3.2
Price
Free
Runs on
Linux, x86_64, glibc
Download
88 MB, as a .deb, an .rpm or one file
Agent
Claude Code on your plan, ChatGPT (preview), or a model of your own
References
dblp search, and a check of every entry
Compiles with
Your TeX Live, or the Tectonic that comes with it
Spelling
US and UK English, LaTeX-aware

Questions

Why not Cursor or VS Code with an agent?

They are built for code. A change to a paragraph comes back as whole lines in a diff, and nothing knows about the PDF. stet shows prose changes word by word and marks them on the typeset page. If you are happy reviewing LaTeX as line diffs, you do not need stet.

How is it different from a web editor with AI, like Overleaf’s or Prism?

Those are web editors: your paper lives on their servers and you use their model, and in return you get live collaboration. stet is an app on your computer that works on your own files with the Claude plan you already have. It has no live collaboration.

What about my co-authors?

stet works on a folder, so share it the way you share code: with git. If the paper is on Overleaf, clone it with Overleaf’s git access, work in stet, and push; your co-authors keep using Overleaf.

Am I allowed to write a paper with AI?

That is between you, your co-authors and your venue. What stet adds is that you see every word the agent wrote and decide on each one, which makes “I checked it” true.

Do I need an API key?

No. Sign in with Claude and the agent runs on the subscription you already have. ChatGPT sign-in is in preview.

Can I use a local model?

Yes. Choose Custom and give the address of anything that speaks OpenAI’s chat API with tool calling: Ollama, LM Studio, llama.cpp, or a hosted service. With a local model nothing of your paper leaves the computer. How well it edits depends on the model.

How do I know a reference is real?

References, Verify looks every entry of your bibliography up in dblp and Crossref. It says which are found, which differ from the record, which cite a preprint of a paper that has since been published, and which are found nowhere.

Where does my paper go?

Nowhere new. Your files stay in their folder on your computer; stet has no server and no account. What the agent reads is sent to the provider you signed in with, exactly as when you paste it into their chat window.

Does it work with my Overleaf project?

stet opens a folder of .tex files. If you have your Overleaf project as a git clone, open that folder.

Do I need to install LaTeX?

No. stet comes with Tectonic, a TeX engine that downloads what a paper needs the first time it compiles it (about 50 MB, so it needs the network once). If you have TeX Live, stet uses that instead, and you can choose per paper in Settings. Build output goes to a cache folder, not into your paper.

Is there a Mac or Windows version?

Not yet. stet is for Linux on x86_64.

Start writing with stet

Same AI. Fewer tabs. It is free, and there is nothing else to install first.

Ubuntu, Debian, Mint

Download, open the file, and choose Install. stet is then in your app menu.

Download .deb

Fedora, openSUSE

Download, open the file, and choose Install. stet is then in your app menu.

Download .rpm

Any other Linux

One file that runs as it is. Allow it to run in its Properties, under Permissions, then open it. It offers to join your app menu.

Download the File
Prefer a terminal?
$ chmod +x stet-x86_64
$ ./stet-x86_64

stet 0.3.2, 88 MB, for x86_64. SHA-256 of the file: 6cdb843ce9fa02795cba53e442be3d496cd987849b89be51dedc6c1c4e597c52

Not on Linux?

Leave your email and you will hear once, when stet runs on your computer.