Scholarcy alternatives that are not Scite or Paperpile

Looking for a real-world alternative to Scholarcy for processing research PDFs. My team's current workflow is brittle: downloading PDFs, manually extracting summaries and references, then copying that data into our CRM and project management tools. Scholarcy's API looks decent for automation, but the pricing is a blocker for our volume.

I need something that can:
* Process academic PDFs and reliably pull structured data (summary, key claims, references).
* Offer a clean API or webhook to push that data into other systems (like Workato or a custom app). No manual exports.
* Handle at least a few hundred papers per month without breaking the bank.

Scite and Paperpile are out (different use cases). I've evaluated a few, but I'm skeptical of claims until I see the integration specs.

Here's what I'm testing, focusing on the integration angle:

**Zotero with plugins (like Zotfile, Better BibTeX)**
* **Pro:** Free, highly customizable. Can set up watched folders for auto-import, then use Zotero's API to fetch metadata.
* **Con:** The summarization is weak. You'll need another layer (like a local LLM script) to get Scholarcy-like summaries. Makes the pipeline more complex.
`// Example of pulling an item via Zotero's API`
`curl "https://api.zotero.org/users/12345/items?format=json&v=3" -H "Zotero-API-Key: your_api_key"`
* **Verdict:** Good foundation, but becomes a middleware project itself.

**Elicit**
* **Pro:** Actually designed for semantic search and summarization of research. Can ask specific questions of a paper.
* **Con:** API is in beta, focused more on their own interface. Getting a clean, automated dump of processed papers into another system feels secondary. Might hit similar pricing walls.

**Local LLM setup (e.g., using PrivateGPT, LlamaParse)**
* **Pro:** Ultimate control. Can parse PDFs, chunk text, and generate summaries entirely in your own environment. No API costs after initial setup.
* **Con:** Significant dev/MLOps overhead. Accuracy varies. You're building the entire pipeline, not integrating a service.

Has anyone actually wired any of these into a production workflow? My primary concern is reducing the manual step, not just having a better reading assistant. The middleware layer needs to be robust.