chat-🗂️ Testing Filesystem Tool.md

USER

Let’s test this new fs tool

ASSISTANT

Sure! Let me start by asking you to share a local folder so I can browse it.

Tool Executed

A folder has been shared! Let me take a look at what's inside:

Tool Executed

Here's what's in the shared **picker-agent** folder:

Name	Type	Size
📁 example_folder/	Directory	—
📄 filesystem_tools.py	File	~20.6 KB
📄 AGENTS.md	File	~4.5 KB

I can now:

• List deeper into example_folder/
• Read the contents of any file
• Search for text patterns across all files

What would you like to explore? 🔍

USER

agents?

ASSISTANT

Let me read the AGENTS.md file for you!

Tool Executed

Here's a summary of the **AGENTS.md** file:

---

📁 picker-agent

A read-only Open WebUI toolkit that gives LLMs access to a user's local filesystem via the browser's File System Access API.

How It Works

• A single Python file (filesystem_tools.py) pasted into Open WebUI's Workspace > Tools
• Python builds JavaScript → sends it to the browser for execution → parses the JSON result back
• Everything lives in the browser tab's memory (no server-side storage)

🔧 Four Tools

Tool	Purpose
open_local_folder	Injects a modal for the user to pick a folder (must be called first)
list_files	Recursive directory listing (max 500 entries)
read_file	Reads file slices with pagination support (default 20KB chunks)
search_files	Regex grep across text files (max 100 matches)

⚠️ Key Details

• Chromium-only (Chrome/Edge 86+) — not supported in Firefox or Safari
• Folder handles are stored per chat ID and lost on tab refresh
• All JS uses try/catch to always return a JSON response (avoids hanging)
• showDirectoryPicker() requires a user gesture (click), hence the modal overlay

🔮 Future Work

• Persist handles via IndexedDB
• Add a file_info metadata tool
• Auto-instruct models about the tool lifecycle

---

Want me to dive into the actual filesystem_tools.py source code or explore the example_folder?