I've had a ChatGPT subscription for a while. It was genuinely useful for a bit, and then it just... wasn't. So I cancelled it (which props to OpenAI for giving me a month for free when I did), and on the way out I built a small tool to help me sort out my conversations (which I never deleted or sorted properly). That tool is tldrgpt.

Pulling the Plug

I've been finding less and less use for it over time, especially after they got rid of 4o.

To be truthful, GPT 5.5 has been producing really good results - solid research, and some of the coding tasks I handed Codex came back genuinely impressive. Before 5.5 I'd mostly ask it about libraries and usage, and Codex was... well, not good - not because of the harness, but because GPT itself just hadn't been quite there for a long time. 5.5 was a real step up. Credit where it's due, it almost made me question whether I actually wanted to cancel.

But it's still nothing I really want to hang on to. It's just too bland.

The other thing I used it for was talking - taking notes for me, suggesting stuff to research, bouncing off an idea, looking for an alternative point of view. The reality is that's been happening less and less and less, especially because of how dumbed down they've made it in the name of "safety." It's "toxic positivity mind virus" as fuck, so I'd rather not, in general. Meh.

And honestly - local AI has finally been getting good. Good enough that I can save that $20 without blinking.

I'm also severely off-put by Sam Altman and some of the things he's said and (allegedly) done. That's a personal gripe, but it absolutely played a factor.

I Didn't Want to Lose My Conversations

So I asked myself: what's the biggest downside to cancelling?

The answer was the conversations. The valuable ones - research and planning, prompt building, the slow refinement of some abstract idea. What to cook for dinner, even. There's real useful stuff buried in there, and OpenAI very conveniently does not give you a quick, easy way to pull it out.

Now, it's good to mention that after cancelling, your convos are still linked to your account, if you choose to keep your account that is.

So I did the obvious thing and asked GPT to research how to export my own conversations. It pointed me at a proper, safe tool for the job - so I didn't have to go and build a scraper myself. Then I specced out a little app to sort through all of it, handed the spec to Codex, and let it implement.

And again - props to GPT 5.5. Even though my prompt was extremely specific, the result was amazing. The result is tldrgpt.

Yes, I used Codex to digest the ChatGPT history I was leaving behind. The irony is not lost on me.

I figured it's only fair if I make it shoot itself.

What is tldrgpt

Well, first you need to export them in HTML with Tampermonkey(this is a browser plugin that injects user specified scripts into web pages) + ChatGPT Exporter.

Q: Why HTML? A: Easy to render in an iframe inside the app and also easy to serialize for analysis.

Then we use tldrgpt - a small local Flask app. You point it at a folder of exported ChatGPT conversations and it:

• scans the folder and lists every conversation in a searchable sidebar
• renders the original conversation so you can actually read it
• serializes each one into clean structured JSON
• runs a local Codex pass over each conversation to generate a TLDR summary + topic tags. The codex pass-over is in 1-shot CLI mode on the backend. Local AI with Ollama, LMStudio, VLLM and LlamaMan is also supported. See the tldrgpt README for setup. If you are using local AI, just make sure you are using a model with a context window that will be able to handle your longest conversation and you're golden. 🪙
• lets you search by title or tag, and soft-delete the junk

The point isn't to keep talking to a model. The point is to turn a pile of opaque chat logs into a searchable, tagged, skimmable archive that lives on your disk - so cancelling costs you nothing.

Setup

Two halves: get your conversations out of ChatGPT, then run the app over them.

Step 1 - Export your conversations

The exporter relies on a userscript, so you need a userscript manager first.

1. Install Tampermonkey in your browser.
2. Install the ChatGPT Exporter userscript from Greasy Fork.
3. Open/reload ChatGPT - the script adds an export button to the UI. Export each conversation you care about as HTML.
4. Drop all those .html files into a single folder. That folder is your source data.

Step 2 - Get the app

git clone https://github.com/nullata/SideQuests.git
cd SideQuests/tldrgpt

Step 3 - Windows (native)

If you've got Python on Windows, run it straight from PowerShell - no WSL needed.

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env

(If PowerShell blocks the activate script, run .venv\Scripts\activate.bat from cmd instead.)

Open .env and point SOURCE_DATA at your exports - use forward slashes to dodge the backslash-escaping headache:

SOURCE_DATA=C:/Users/you/chatgpt-exports

Then run it and open http://127.0.0.1:5000:

flask --app app run --debug

Step 4 - Linux / macOS / WSL

Same on all three.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Set SOURCE_DATA in .env to wherever you dumped the exports, then run it and open http://127.0.0.1:5000:

flask --app app run --debug

If you're on WSL and the exports live on the Windows side, your drives are mounted under /mnt/c/..., so C:\Users\you\chatgpt-exports becomes:

SOURCE_DATA=/mnt/c/Users/you/chatgpt-exports

Analysis

To analyze a conversation you need an AI backend - either Codex or local AI. Everything else (browsing, serializing, viewing the raw JSON, soft-deleting) works without one; only the TLDR summary + topic tags step calls out to a model. You pick the backend in .env with ANALYSIS_BACKEND:

• codex - shells out to the local Codex CLI via codex exec in one-shot mode. Needs Codex installed and logged in; model set with CODEX_MODEL.
• openai - points at any OpenAI-compatible /v1 endpoint (Ollama, LM Studio, vLLM, LlamaMan). Set OPENAI_BASE_URL and OPENAI_MODEL.

You can analyze convos individually or in bulk. Just make sure to monitor your usage if working with Codex. The rest of the UI actions are pretty easily understandable once you look around for a second.

Deleted a conversation on accident? No worries - as stated before - the app has a soft delete. So it will place "deleted" conversations in a directory of your choosing, which you can later get rid of whenever you like.

Conclusion

I'm not anti-AI - I use it every day. It's that I just don't want to rent my own thinking back from a company I've kind of soured on, mostly (among other things) for enshitifying their product. Especially when the useful stuff of all those convos is mine and local AI is finally good enough to stand in, no thanks to anything open from OpenAI anyway.

So: export it, archive it, tag it, own it. Then cancel and be GPT-free.

via GIPHY

• GitHub: github.com/nullata/SideQuests

Live long and prosper. 🖖👽