Agent in Your Filesystem. No Installation Needed?

I remember Anthropic releasing Opus 4.8 a couple of months ago and people on Twitter making fun of its inability to tell how many days of the week have the letter d in them. I actually had trouble believing that, thinking these issues would be fixed by now. But no, that was still the case on claude.ai chat: it confidently proclaimed the answer was two—Sunday and Monday! I get how LLMs view words as combinations of tokens and the difficulty of inspecting individual letters, but surely it could reach for Linux tools like grep to count characters, right?

Claude Code on a user’s machine can obviously invoke CLI tools, so why not do that in the chat interface too? Well, we are moving in the direction where they reason, use external tools and connections, spawn tasks, and in general become more agentic. It just hasn’t been happening fast enough, especially outside the developer community.

I think the main challenge is bringing all of this to environments where users actually are. Not desktop apps, but the web and mobile. Claude Chat and ChatGPT are increasingly using scripts, isolated execution environments, and web search to augment the pure-LLM approach. None of these individually are particularly impressive; it’s just that bringing that seamless experience to, for example, the web has not been trivial.

I’ve been thinking about what’s still missing from browsers to make web-based AIs fully agentic. Both to enable regular users to use AI in more advanced ways and to enable power users to use it in more casual environments. What’s currently preventing me is working with files—probably the only reason I still use desktop AI apps. For someone else, it might be a different reason, but it’s almost always related to the web lacking a feature; something that makes the AI less agentic. It’s for that reason, it’s probably useful to first agree on what agents even are.

Agentic qualities

How is a chat interface (like Claude Chat or ChatGPT) fundamentally different from an agent interface (like Claude Code, Codex, or OpenCode)? Anthropic’s own attempts to distinguish between different paradigms seem overly convoluted—these categories are clearly being erased. I hate the word, but it really is a spectrum. Either way, the definition is in the name: a system is more agentic if it has more independence or agency.

A senior employee clearly has more agency than a junior one, so what separates them?

• knowledge and skills
• ability to use tools and delegate
• ability to infer from available context¹

We can view AI agents through the same lens.

Knowledge

As a rough approximation, the larger the base model, the more knowledgeable it is. More knowledge means a system can answer a greater range of questions, but it also helps with the other two qualities of delegation and context inference.

Delegation

High agency is knowing when to delegate to tools or less powerful agents. This expands the domain of tasks that can be accomplished. Or accomplished without going bankrupt.

Depending on the level of agency, asking “What day of the week is Christmas 2028?” could give

hallucinated answer

Tuesday

expensive reasoning chain that might give the correct answer

Reasoning

I'm working through the day of the week for December 25, 2028 by anchoring to known dates. I know December 25, 2025 falls on Thursday, and I'm using that as a reference point to calculate forward, accounting for leap years and the number of days that pass between these dates. I need to recalculate the day-of-week progression more carefully. Starting from 2025 Jan1 as Wednesday, I'm adding the number of days in each year modulo 7: 2026 Jan1 is Thursday (365 mod 7 = 1), 2027 Jan1 is Friday, 2028 Jan1 is Saturday, and 2029 Jan1 is Monday (accounting for 2028 being a leap year with 366 days, which is 2 mod 7). For Dec25 2028, that's day 360 of the leap year, meaning 359 days after Jan1, so I need to find what day Saturday plus 359 days lands on. 359 mod 7 equals 2, so Saturday plus 2 days gives Monday. Let me verify this against a known reference point: 2025 Dec25 is Thursday, and working backward from 2024 Jan1 being Monday confirms the chain is correct, so Christmas 2028 falls on Monday.

Monday

high-confidence answer using a cheap tool call

Tool call

python3 -c "import datetime; print(datetime.date(2028, 12, 25).strftime('%A'))"

Monday

Context

Tools can expand the available context, while a more powerful model can infer more from less. But ultimately, the information available is the one thing that agents still depend heavily on users for. How does it compose an email reply to your boss? You either paste the email or just connect your Gmail. How does it complete a Python function for verifying user passwords? You either paste the SQL schema or just give the agent access to your whole codebase.

Pasting stuff in is easy, but you have to do it every time. Not only that, you are also limiting the context to the one snippet you’ve provided—which makes it hard for the model to infer your style, habits, and additional information that you hadn’t considered. And yet, that’s how most people interact with AI—through chat, pasting things in. It’s a massive problem and it’s why I became interested in whether browsers are really limiting us and if there’s a solution.

Browsers have spoiled us

For many people, the browser is the only program they use. I could install some obscure Linux distro on my grandma’s computer, but as long as there was a big Chrome logo somewhere, she would be fine. Knowledge workers can run an email client, a word processor, Slack, Figma, and yes, AI, in their browser. If they can avoid installing anything else, they will.

However, for some things the browser just isn’t enough. You can chat to an AI on the web, you can connect it to your email and Slack, you can even give it a sandbox to execute commands. But a web-based AI can’t manipulate the files on your computer. Right?

File System API

I stumbled across this API accidentally while browsing MDN (it’s a fun exercise to do from time to time, just to realize how much the browsers can do without any additional libraries). The File System API enables websites to read and write to user-approved folders:

• for await (const handle of dir.values()): list files
• await handle.getFile(): get file
• await writable.write("hello"): write to file

People have built some interesting prototypes with it. Ben Holmes made a website that edits your Markdown notes locally, while the VS Code team shipped their IDE on the web, allowing you to edit files locally from within the browser.

Agent that cleans your folder

To me, the potential of an agent in a browser being able to access the filesystem is obvious. It brings the one missing piece of context that an agent needs to an environment where a normie user feels familiar. We need more demonstrations of the value agents provide, and doing that without the user needing to install anything is perfect. Amrit Baveja built an impressive system where agents can perform arbitrary actions on the files using efficient traversing and indexing. I wanted to build a simpler demo, something that is dead easy to explain to anyone, i.e., doesn’t require you to think of a problem to tackle.

My idea is an agent that cleans your messy folders. It’s a bit cliché, but it’s definitely a problem I still have. The files that accumulate in Desktop and Downloads folders are a complete mess—there are a lot of them, and you’re unsure which ones can be safely deleted. An AI agent would read the directory, maybe extract some text and image content and rename the files and/or put them into folders. In fact, I downloaded Claude Code to my grandma’s computer a few weeks ago just for this reason. This time, I wanted to see if I could make it work in a browser. You can check out the demo website AI Janitor! The gist of it is shown below:

It works! But are there any problems? For sure…

• only Chromium supports this and only on desktop—having worked on podcasting tech the past few years, I’m well aware you shouldn’t bet your business on big tech adopting new standards quickly
• Chrome prevents selecting Desktop, Downloads, or Documents, so the aforementioned demo is not as straightforward—you have to choose a subdirectory of those
• the current Chrome implementation could be improved, e.g., adding directory renames would avoid inefficient pasting and deleting of files
• security is obviously a concern, though the only difference from a desktop app is that a browser experience would have less friction

It’s obviously not ready, so why be optimistic about browsers? Because it’s the most amazing medium, and if something can eventually be done via the browser, it will be. Even if File System API takes time to mature, any limitations browsers have in the meantime will get worked around (with cloud storage, for example). That’s much more likely than people moving to desktop apps. Whatever the path, I’m incredibly excited for a more intelligent future: language models using tools and having the right context, working with humans where they feel most comfortable.