Crate

Drag in a folder. Ask your agent. It just works.

Crate turns any folder of files — PDFs, code, notes, CSVs, screenshots — into an agent-native, searchable, citable workspace with a single drag-and-drop. Files are indexed entirely in your browser (nothing is ever uploaded), and Crate registers a set of WebMCP tools that any AI agent visiting the page can call: ChatGPT's in-app browser, Chrome with the WebMCP flag, the Nekuda Workbench, Stagehand, and more.

The human sees a beautiful reader. The agent sees typed tools. They share the same session, the same documents, the same cursor.

Routes: / is the landing page; /app is the workspace.

Why WebMCP is load-bearing

Without WebMCP, an agent would have to scrape the DOM or OCR screenshots to work with your files — slow, fragile, hallucinated. With WebMCP, the tools are declared via document.modelContext.registerTool; the agent calls them; the page executes them on the user's actual local files with full session state. Every call is visible in the on-page agent activity strip, and cite_source draws a live citation ring around the exact span the answer came from.

The tools (8 imperative + 2 declarative)

Tool	Kind	What it does
get_crate_status	imperative	Always-on status check: document count, kinds, which tools are live
search_documents	imperative	Hybrid BM25 + semantic search, returns ranked snippets with scores
open_document	imperative	Full text of a doc (or one PDF page); opens it in the viewer
cite_source	imperative	Rings a quoted span in the viewer so the user sees the anchor
summarize_doc	imperative, confirm: true	Writes a 3-bullet extractive summary to the doc's panel
compare_documents	imperative	Side-by-side table for 2–5 docs, rendered on the page
extract_table	imperative	Pulls tabular data out of CSVs / Markdown tables / PDFs as CSV
export_table	imperative, confirm: true	Triggers a real CSV download in the user's browser
ask_my_files	imperative	One-shot Q&A composing search + cite with inline citations
<form toolname="ask">	declarative	The persistent ask box (browser synthesizes the schema)
<form toolname="upload_folder">	declarative	Native file-picker fallback for agents

Registration is state-aware: with an empty Crate only get_crate_status exists; the full toolset appears when files are indexed and toolchange tells agents to re-read the list. All handlers honor the agent's AbortSignal, use annotations per the spec (readOnlyHint for reads, untrustedContentHint on outputs that echo file content, a confirm schema field for anything that writes to disk), and log visible activity events.

Privacy: where the data lives

Nowhere but your browser. Extracted text and embeddings go into IndexedDB; nothing is sent to any server — there is no server API at all. Close the tab, come back tomorrow, your Crate is still there. Delete it with one click.

Search

• Lexical: BM25 over ~900-char chunks of extracted text
• Semantic: all-MiniLM-L6-v2 via transformers.js, WebGPU-accelerated when available, WASM fallback — running in your browser, no API key
• Hybrid: 0.55 × BM25 + 0.45 × cosine, gracefully degrading to lexical-only if the model can't load
• OCR: image files (screenshots, photos of whiteboards, scanned docs) are read with tesseract.js — also fully in-browser

File types

PDF (pdf.js, page-by-page), Markdown, CSV, code (30+ extensions), plain text, images — OCR'd locally with tesseract.js (WASM) so screenshots become searchable. Files over 25 MB are skipped. node_modules, .git, and friends are ignored when you drop a folder.

Run it

npm install
npm run dev        # http://localhost:3000 → landing at /, workspace at /app
npm test           # 64 unit tests (vitest): chunking, search ranking, citation
                   # matching, table extraction, filestore, WebMCP tool handlers

Or with Docker:

docker build -t crate . && docker run -p 3000:3000 crate

To see the agent side, open the URL in an environment where WebMCP is enabled:

• ChatGPT's in-app browser (easiest) — ask "what's in my crate?"
• Chrome 149+ with chrome://flags/#enable-webmcp-testing → Enabled
• Nekuda WebMCP Workbench extension — inspect and call tools manually
• Model Context Tool Inspector extension — natural-language agent for testing

Without WebMCP, everything except the agent-facing tool registry still works: drop files, search, ask, read — the badge in the header tells you WebMCP is off and how to enable it.

Testing notes

• Install deps, run dev, drop a folder.
• First semantic search downloads the ~25 MB model once (cached afterwards).
• Try: "find every doc that mentions revenue and quote the number" → the agent should run search_documents → open_document → cite_source, and you'll see the ring animate.

Stack

Next.js (App Router) · TypeScript · idb (IndexedDB) · pdf.js · transformers.js · papaparse · hand-rolled CSS (paper/ink/green, Space Grotesk + JetBrains Mono). No backend, no auth, no API keys.

License

MIT — see LICENSE.