Open Crate
WebMCP native · OpenAI WebMCP Challenge

Drag in a folder.
Ask your agent.

Crate turns any folder of files into a searchable, citable workspace for you and your AI agent. Everything runs in your browser. Nothing is ever uploaded.

No account · No API key · No upload · MIT licensed
Privacy

Your files never leave your browser

There is no server API. Text extraction, OCR, embeddings and search all run inside the tab. Your data sits in IndexedDB on your machine and is gone when you clear it.

The agent side

11 declared tools

search_documentsopen_documentcite_sourcesummarize_doccompare_documentsextract_tableexport_tableask_my_filesget_crate_statusform toolname="ask"form toolname="upload_folder"
Trust

Answers with proof

cite_source draws a green ring around the exact span the answer came from, live in the viewer, while you watch.

Search

Hybrid by default

BM25 keywords plus MiniLM embeddings on WebGPU, degrading gracefully to lexical only when the model is unavailable.

Images

Screenshots become searchable

Every image is OCR'd locally with tesseract.js. That folder of whiteboard photos is now a queryable corpus.

Agents

Works with your agent

ChatGPTChromeNekuda WorkbenchStagehandClaude

Same URL, same session, same files. The agent inherits your browser.

Safety

Reads are read. Writes confirm.

Every tool carries spec annotations. The two mutating tools require the agent to ask you first, then pass confirm.

How it works

1

Drop

Drag any folder onto the page. Crate walks it, skips the junk, reads every file locally.

2

Index

Text is extracted, OCR runs, everything is chunked and embedded in your browser. The toolset goes live.

3

Ask

Your agent searches, opens, cites and exports. You watch every call in the activity strip.

Your files were always yours. Now they speak agent.

Questions judges ask

Do my files get uploaded anywhere?

No. There is no upload endpoint in the codebase. Text extraction, OCR, embeddings and search all run inside your browser tab. Data is stored in your browser's IndexedDB and stays on your machine.

What do I need to make the agent work?

Open the site in ChatGPT's in-app browser, or Chrome 149+ with the WebMCP testing flag enabled, or any agent that speaks WebMCP. Without one of those, everything except agent invocation still works: drop, search, ask, read.

Which file types does Crate understand?

PDFs (page by page), images (OCR'd locally, so screenshots become searchable), Markdown, CSV, code in 30+ languages, and plain text. Files over 25 MB are skipped.

What does WebMCP change compared to screen scraping?

The agent calls typed functions with JSON schemas instead of guessing at pixels and DOM. It is faster, more reliable, and runs on your session with your files, while you watch every call happen on the page.

Open the live app. Drop a folder. Ask.

Open Crate →