If you've used ChatGPT for a year, your sidebar is a graveyard. Old debugging sessions, half-finished ideas, conversations you'd rather not keep. And OpenAI gives you exactly one way to deal with them: hover a row, open the menu, click delete, confirm. One at a time. Forever.
There are extensions that fix this. I didn't install any of them.
Not because they're malicious — I have no evidence of that. Because of what they ask for. An extension that manages your chats needs to run on chatgpt.com with your session active. That is the same page that holds every conversation you've ever had. If the code is minified and the permissions are broad, "trust me" is the only thing on offer.
So I built my own, and made it boring on purpose.
Bulk Delete for ChatGPT is now live on Mozilla's add-on site: addons.mozilla.org/en-US/firefox/addon/bulk-delete-for-chatgpt
It is unofficial, MIT licensed, and at the time of writing has exactly zero users. That's the honest state of it.
The constraint I started with
The feature is trivial. GET the conversation list, DELETE the ones selected. That's maybe 40 lines.
The hard part is everything around it:
- How do you prove the extension can't delete something you didn't tick?
- How do you stop a misclick from wiping a year of history?
- What happens when the session expires halfway through 200 deletions?
- How do you know the deletes actually happened, instead of the UI just saying so?
Those questions produced most of the design.
Six decisions that matter
1. It is a dry run until you say otherwise
The first button press does nothing. It walks the whole flow, logs what it would delete, and changes no state.
You have to press a second time, in a different mode, to delete anything. That single extra step has caught every accidental click I've made while testing.
2. There is no "delete all"
No "select all", no "delete filtered", no date rules. The codebase contains one deletion loop, and it iterates over a Set of IDs that came from checkboxes.
"Select all shown" exists, but it only applies to the current search filter and it tells you the count before you commit — Select all 47 shown. If the number looks wrong, you stop.
That distinction matters. A "delete everything older than 30 days" button is one bad regex away from a very bad afternoon.
3. The token is never stored
The extension doesn't ask you to paste an API key, and it doesn't have a login screen. At the start of each run it reads the session token from the page's own /api/auth/session endpoint, keeps it in memory for the duration of the run, and lets it go.
Nothing is written to localStorage, chrome.storage, or a cookie. There is no server to send it to — there is no server at all.
4. It stops on the errors that mean stop
Auth failures and rate limits are not retried. They abort the run and tell you what happened:
if (res.status === 401) throw unauthorized();
if (res.status === 429) throw rateLimited();
Transient failures get exactly one retry. A 404 on that retry is treated as success — the chat is already gone, which is the outcome you wanted:
if (res.status === 404 && attempt === 1) {
return { ok: true, status: 404, alreadyGone: true };
}
This sounds obvious until you've watched a naive script hammer a rate-limited API for 300 requests.
5. It verifies the result, not the promises
After a run, the extension re-reads the server's own conversation total and compares:
Server total 312 → 300 (Δ12)
If 12 chats were reported deleted, the total must have dropped by 12. If it didn't, you see that immediately instead of assuming it worked.
6. You type the number to confirm
The final confirmation requires typing the count of selected chats. Not clicking "OK" — typing 12.
It's slightly annoying. That's the point. It's the difference between a misclick and a decision.
The test I actually care about
Unit tests would prove the loop counts correctly. The thing I wanted to prove was different: that only ticked conversations get deleted.
So the repo has an end-to-end suite that runs the real extension in a real browser against a fake chatgpt.com — a local server that mimics the session endpoint, the list endpoint, and the delete endpoint, behind a self-signed certificate. It ticks a subset, runs a delete, and asserts on the server side that the deleted IDs are exactly the ticked IDs.
Every safety claim above is a test, not a comment.
Install
Firefox gets a one-click install:
Open the page, press Add to Firefox, then Add. Reload chatgpt.com. There's a small "Bulk delete" button in the bottom-right corner.
Chrome, Edge, and Brave work too, but by hand — the Chrome Web Store charges a one-time developer fee and I haven't paid it. The releases page has the zip; chrome://extensions with Developer mode on will load it.
Back up first. ChatGPT → Settings → Data controls → Export data. Deleting a conversation is permanent, the extension has no undo, and the export is the only way back.
The limitations, honestly
- It depends on ChatGPT's private endpoints. If OpenAI changes them, it breaks. It's a small codebase, but it's still a moving target.
- Firefox on Android is untested. I deliberately left it out of the listing rather than claim support I haven't verified.
- No undo. By design — a "recently deleted" bin would mean the extension keeps a copy of your chats, which defeats the point.
- Zero users. This is a v1.0.0 that just passed review. If something is broken, an issue on GitHub is the fastest way to get it fixed.
Why publish it at all
Because the alternative was another minified extension asking for broad permissions on a page where people keep their most private conversations. The source is four files and about a thousand lines, with no build step. You can read all of it in one sitting and decide for yourself.
That's the whole pitch. Not "fastest", not "smartest". Small enough to check.
Top comments (0)