Library or hosted? Choosing between edgartools and edgar.tools
If you found this blog, you probably already pip install edgartools. About a million people a month do. So the question is not "what is edgar.tools" — it is whether you need it, given that the library is free and you already know it.
Short answer: they are the same parsing, in two places. The library runs on your machine and hands you data. The hosted platform runs the same discipline on our machines and hands the answer to your assistant. Which one you want depends on what you are doing at the time.
Stay in the library when you are working the wire
The library reads a filing the minute it lands on EDGAR. When the data does something strange — a manager moves its 13F book to a parent entity, a filer stuffs three exhibits into one document — you are holding the steering wheel. You can read the notice yourself, follow it to the sibling CIK, and build whatever baseline you decide is true.
That is library territory, and it always will be:
- Bulk pipelines and backfills across thousands of filings
- Custom analytics where you own the join logic
- Anything that feeds a model you are training
- Notebooks, where the point is to look
Nothing about the platform changes this. The library stays MIT-licensed and free.
Go hosted when you want the answer, not the script
The other half of the week you are not writing a pipeline. You want to know what changed in a portfolio last quarter, whether the CEO sold in the window, or what a 10-K says about leases this year that it did not say last year. You want the reconciled view, and you want to ask the follow-up in one sentence.
That is what the hosted platform does. Your assistant — Claude, ChatGPT, Cursor, or your own agent loop — connects to edgar.tools over MCP and asks questions in plain English. What comes back is structured data with its source filing attached: a manager resolved from a name, portfolio weights already computed, a disclosure diff already run, a freshness caveat stated instead of hidden.
Three things you get that the library does not give you:
- The reconciled view. Filer restructurings, amendments, and consolidation lag are absorbed before the answer reaches you. The 13F ritual that lies once a year stops lying.
- Provenance in the answer. Every figure links to the filing it came from. Your assistant is told to cite, not to paste.
- No editor required. The workflow exists when you are on a phone, in a meeting, or three questions past where the script stopped.
Connecting takes about a minute: paste one address into your assistant's connector settings and approve the sign-in. The free tier is enough to run the examples on this blog. The Pro trial opens the full research toolkit — financials, filing search, insider activity, ownership, material events — with a card on file and nothing charged until the trial ends.
Or connect the free tier · see what each tier includes
The honest split, in one table
| You are… | Use |
|---|---|
| Building a dataset or a backtest | Library |
| Reading Friday's filing on Friday | Library |
| Deciding what "previous quarter" means for a filer that restructured | Either |
| Asking "what changed?" and then "just the Amazon line" | Hosted |
| Answering a question away from your editor | Hosted |
| Handing SEC research to an agent that cannot run Python | Hosted |
Most people who use both settle into the same rhythm: the library for the work that produces something, the platform for the questions that come up while doing everything else.
See it run both ways
Every post in the your edgartools workflow, now in Claude series takes one real Python ritual and runs it both ways, wrong turns included:
- Do your 13F workflow in Claude — the quarterly diff that lied, and the one-sentence version
More coming. If there is a workflow you want treated this way, open an issue and say so.
I'm Dwight Gunning. I wrote the library, and the platform is my own continuation of that work: the same parsing, packaged as a data layer your assistant can query. Your ritual script does not retire. It gets a second front end.