# 2 GB user interviews into Claude when the chat limit is 500 MB

> Source: <https://blog.zernote.com/2gb-user-interviews-into-claude/>
> Published: 2026-07-31 14:38:50+00:00

**Two months ago we moved Zernote's interface into MCP.**

We ran **37 interviews** with UX researchers. Around **80%** were already running **interview analysis** through an **LLM**. They paste transcripts into Claude, ask for themes. After **7 to 9 interviews** the **model** starts to **hallucinate**.

**So we built the wrapper.** Tools for folders, interviews, transcripts, quotes. OAuth. Claude could read a corpus and cite it.

And we still could not turn the UI off.

One thing blocked it: **upload limits**.

User researchers often record calls as **screencasts**. Such recordings can be up to 800 MB, 900 MB, and sometimes **1800 MB**.

Claude has a hard limit: **500 MB** per file.

We went looking for how other people were getting files into Claude at all.

**Brain Graph's** connector was doing it. Their tool takes a PDF, slides, or an audio file and lands it in their knowledge graph. Their caps are modest: around **150 MB** for audio, **35 MB** for documents. **Nowhere near a 1800 MB** screencast.

The limits were not the interesting part. **The mechanism was.**

They were not pushing bytes through the tool call. **They were opening a widget.**

Anthropic [shipped MCP Apps](https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/) in January 2026. Any connector can hand Claude an iframe, **a small application rendered directly inside the conversation**.

The iframe does not run on claude.ai. Anthropic serves it from its own sandboxed origin:

```
https://<hash>.claudemcpcontent.com
```

Your widget talks to your server through that origin.

Which makes the path: **user → browser → proxy → your backend**.

**Bingo.**

The file goes straight from the browser to our backend. It never enters the context window.

The next step was **signed URLs to the bucket**.

Now the upload bypasses our backend entirely. The **file goes straight** from the browser **to storage**. The widget notifies the backend afterwards, once the file has landed.

After the upload, the file-processing microservice checks whether the user has enough tokens to process the file. If not, it returns a 403.

At the end of the upload the widget can insert a full prompt into the chat: the IDs of the uploaded files and the next steps.

The prompt is in English, with a request at the end to reply in the user's own language.

The last detail, and not a small one: presentation.

The widget renders inside the chat as an HTML block. Without careful styling its **interface hurts to look** at and feels foreign to the conversation.

We fixed that.

*We got obsessed with working with our own tools through MCP and widgets.*

*The images in the article you are reading were uploaded through one.*

This version of the MCP is already live in [our connector](https://zernote.com/mcp). It is free to try.

If you have any questions - feel free to text me at [ceo@zernote.com](mailto:ceo@zernote.com) :)