I use Playwright to intercept all requests and responses and have Claude Code navigate to a website like YouTube and click and interact with all the elements and inputs while recording all the requests and responses associated with each interaction. Then it creates a detailed strongly typed API to interact with any website using the underlying API.
Yes, I know it likely breaks everybody's terms of service but at the same time I'm not loading gigabytes of ads, images, markup, to accomplish things.
If anyone is interested I can take some time and publish it this week.
I also do this. My primary use case is for reproducing page layout and styling at any given tree in the dom. So, capturing various states of a component etc.
I also use it to automatically retrieve page responsiveness behavior in complex web apps. It uses playwright to adjust the width and monitor entire trees for exact changes which it writes structured data that includes the complete cascade of styles relevant with screenshots to support the snapshots.
There are tools you can buy that let you do this kind of inspection manually, but they are designed for humans.
---
My first reaction to seeing this FP was why are people still releasing MCPs? So far I've managed to completely avoid that hype loop and went straight to building custom CLIs even before skills were a thing.
I think people are still not realizing the power and efficiency of direct access to things you want and skills to guide the AI in using the access effectively.
Maybe I'm missing something in this particular use case?
Yes please, maybe there will be some solution that will fit the problem better! I recently released something similar, and because of the small API, I'm more comfortable using it.
Exactly, it is an agent skill that interacts pressing buttons and stuff with a webpage capturing and documenting all the API requests the page makes using Playwright's request / response interception methods. It creates and strongly typed well documented API at the end.
Sounds awesome. I've been using mitmproxy's --mode local to intercept with a separate skill to read flow files dumped from it, but interactive is even better.
yt-dlp is relatively stable, but still occasionally breaks for long periods. I get the sense YouTube is becoming increasingly adversarial to yt-dlp as well.
I don't know the details, but it doesn't seem like yt-dlp is running the entire YouTube JS+DOM environment. Something like a real headless browser seems like it would break less often, but be much heavier weight. And Youtube might have all sorts of other mitigations against this approach.
It would probably help people who want to go to a concert and have a chance to beat the scalpers cornering the market on an event in 30 seconds hitting the marketplace services with 20,000 requests.
I can try to see if can bypass yt-dlp. But that is always a cat and mouse game.
To clarify - yt-dlp is a command line tool for downloading youtube videos, but it's in a constant arms race with the youtube website because they are constantly changing things in a way that blocks yt-dlp.
100% I'll response to this by Friday with link to Github.
I use Patchright + Ghostery and I have a cleaver tool that uses web sockets to pass 1 second interval screenshots to the a dashboard and pointer / keyboard events to the server which allow interacting with websites so that a user can create authentication that is stored in the chrome user profile with all the cookies, history, local storage, ect.. in the cloud on a server.
Can you list some websites that don't require subscription that you would like to me to test against? I used this for Robinhood and I think Linked in would be a good example for people to use.
Google is so far behind agentic cli coding. Gemini CLI is awful. So bad in fact that it’s clear none of their team use it. Also MCP is very obviously dead, as any of us doing heavy agentic coding know. Why permanently sacrifice that chunk of your context window when you can just use CLI tools which are also faster and more flexible and many are already trained in. Playwright with headless Chromium or headed chrome is what anyone serious is using and we get all the dev and inspection tools already. And it works perfectly. This only has appeal to those starting out and confused into thinking this is the way. The answer is almost never MCP.
> Also MCP is very obviously dead, as any of us doing heavy agentic coding know.
As someone that does heavy agentic coding (using basically all the tools), this is so far from the truth. People claiming this have probably never worked in large enterprise environments where things like authentication, RBAC, rate limiting, abuse detection, centralized management/updates/ops, etc. are a huge part of the development and deployment workflow.
In these situations you can't just use skills and cli tools without a gigantic amount of retooling and increased operational and security complexity. MCP is really useful here, and allows centralized eng and ops teams to manage their services in a way that aligns with the organizations overall posture, policies, and infrastructure.
> Google is so far behind agentic cli coding. Gemini CLI is awful.
This part I totally agree. It's really hard to express how bad it is (and it's really disappointing.)
Some people will push back on this. They are holding out hope that the recent improvements Anthropic has made in this regard have improved the context rot problem with MCP. Anthropic's changes improve things a little. But it is akin to putting lipstick on a pig. It helps, but not much.
The reason MCP is dying/dead is because MCP servers, once configured, bloat up context even when they are not being used. Why would anybody want that?
Use agent skills. And say goodbye to MCP. We need to move on from MCP.
MCP is very much not dead. centralized remote MCP servers are incredibly useful. also bespoke CLIs still require guidance for models to use effectively, so it's clear that token efficiency is still an issue regardless.
I see remote MCP servers as a great interface to consume api responses. The idea that you essentially make your apis easily available to agents to bring in relevant context is a powerful one.
When folks say MCP is dead, I don't get it. What other alternatives exist in place of MCP? Arbitrary code via curl/sdks to call a remote endpoint?
Tbh I find self-documenting CLIs (e.g. with a `--help` flag, and printing correct usage examples when LLMs make things up) plus a skill that's auto invoked to be pretty reliable. CLIs can do OAuth dances too just fine.
MCP's remaining moats I think are:
- No-install product integrations (just paste in mcp config into app)
- Non-developer end users / no shell needed (no terminal)
- Multi-tenant auth (many users, dynamic OAuth)
- Security sandboxing (restrict what agents can do), credential sandboxing (agents never see secrets)
Imagine if, in addition to local MCP "servers", the MCP people had nurtured a structured CLI-based --help-equivalent consumable by LLMs and shell completion engines alike. Doing so, you unify "CLI" (trivial deployment; human accessibility) and MCP-style (structured and discoverable tool calling) in a single DWIM artifact.
But since when has this industry done the right thing informed by wisdom and hindsight?
I think cli’s are more token efficient- the help menu is loaded only when needed, and the output is trivially pipe able to grep or jq to filter out what the model actually wants
I don't know if this just anecdotal random impression, but in a last week or two I had mostly good experience with Google cli. While previously I constantly complained about it. I have been using it together with codex, and I would not say that one is much better than another.
It is hard to say nowadays, when things change so quickly
I know it’s a bit of a tangent but man you’re right re. Gemini CLI. It’s woefully bad, barely works. Maybe because I was a “free” user trying it out at the time, but it was such a bad experience it turned me off subscribing to whatever their coding plan is called today.
it's not the CLI, it's the model. The model wasn't trained to do that kind of work, was trained to do one shot coding, not sustained back and forth until it gets it right like Claude and ChatGPT.
> Why permanently sacrifice that chunk of your context window when you can just use CLI tools which are also faster and more flexible and many are already trained in
What about all the CLI tools not baked into the model's priors?
Every time someone says "extensibility mechanism X is dead!", I think "Well, I guess that guy isn't doing anything that needs to extend the statistical average of 2010s-era Reddit"
For example, I use codex to manage a local music library, and it was able to use the skill to open a YT Music tab in my browser, search for each album, and get the URL to pass to yt-dlp.
Do note that it only works for Chrome browsers rn, so you have to edit the script to point to a different Chromium browser's binary (e.g. I use Helium) but it's simple enough
On one hand, cool demo, on the other, this is horrifying in more ways than I can begin to describe. You're literally one prompt injection away from someone having unlimited access to all of your everything.
Not the person you're replying to, but: I just use a separate, dedicated Chrome profile that isn't logged into anything except what I'm working on. Then I keep the persistence, but without commingling in a way that dramatically increases the risk.
edit: upon rereading, I now realize the (different) prompt injection risk you were calling out re: the handoff to yt-dlp. Separate profiles won't save you from that, though there are other approaches.
Even without the bash escape risk (which can be mitigated with the various ways of only allowing yt-dlp to be executed), YT Music is a paid service gated behind a Google account, with associated payment method. Even just stealing the auth cookie is pretty serious in terms of damage it could do.
Agreed. I wouldn't cut loose an agent that's at risk of prompt injection w/ unscoped access to my primary Google account.
But if I understood the original commenter's use case, they're just searching YT Music to get the URL to a given song. This appears[0] to work fine without being logged in. So you could parameterize or wrap the call to yt-dlp and only have your cookie jar usable there.
Oh, that's true, even allows you to play without an account. I can swear that at some point it flat out refused any use unless you're logged in with an account that has YT Music (I remember having to go to regular YouTube to get the same song to send it to someone who didn't have it).
As long as it’s gated and not turned on by default, it’s all good. They could also add a warning/sanity check similar to “allow pasting” in the console.
Relying on warnings or opt-ins for something with this blast radius is security theater more than protection. The cleverest malware barely waits for you to click OK before making itself at home, so that checkbox is a speed bump on a highway.
Chrome's 'allow pasting' gets ignored reflexively by most users anyway. If this agent can touch DevTools the attack surface expands far faster than most people realize or will ever audit.
You get used to it :) And especially once you get used to the YOLO lifestyle, you end up realizing that practically any form of security is entirely worthless when you're dealing with a 200 IQ brainwashed robot hacker.
For now you are. All these things fall with time, of course. You will stop caring once you start feeling safe, we all do.
Also. AAarrgh, my new thing to be annoyed at is AI drivel written slop.
"No browser automation framework, no separate browser instance, no re-login."
Oh really, nice. No separate computer either? No separate power station, no house, no star wars? No something else we didn't ask for? Just one a toggle and you go? Whoaaaaaa.
Edit: lol even the skill itself is vibe coded:
Lightweight Chrome DevTools Protocol CLI. Connects directly via WebSocket — no Puppeteer, works with 100+ tabs, instant connection.
I feel like there's nothing fucking left on the internet anymore that is not some mean of whatever the LLM is trained to talk like now.
What can you do? I mentioned the use of AI on another thread, asking essentially the same question. The comment was flagged, presumably as off topic. Fair enough, I guess. But about 80% (maybe more) of posted blogs etc that I see on HN now have very obvious signs of AI. Comments do too. I hate it. If I want to see what Claude thinks I can ask it.
HN is becoming close to unusable, and this isn’t like the previous times where people say it’s like reddit or something. It is inundated with bot spam, it just happens the bot spam is sufficiently engaging and well-written that it is really hard to address.
To be clear, this isn't a skill for the devtools mcp, but an independent project. It doesn't look bad, but obviously browser automation + agents is a very busy space with lots of parallel efforts.
DevTools MCP and its new CLI are maintained by the team behind Chrome DevTools & Puppeteer and it certainly has a more comprehensive feature set. I'd expect it to be more reliable, but.. hey open source competition breeds innovation and I love that. :)
(I used to work on the DevTools team. And I still do, too)
The real problem this thread exposes is that we're duct-taping browser automation (Playwright, CDP, MCP wrappers) onto an interface designed for humans — the DOM. Every approach discussed here is fighting the same battle: too many tokens to represent page state, flaky selectors, hallucinated DOM structures, massive context cost.
What we actually need is a standard for websites to expose a machine-readable interaction layer alongside the human one. Something like robots.txt but for agent capabilities — declaring available actions, their parameters, authentication requirements, and response schemas. Not scraping the DOM and hoping the AI figures out which button to click.
The web already went through this evolution once: we went from screen-scraping HTML to structured APIs. Now we're regressing back to scraping because agents need to interact with sites that only have human interfaces. A lightweight standard — call it agents.json or whatever — where sites declare "here are the actions you can take, here are the endpoints, here's the auth flow" would eliminate 90% of the token waste, security concerns, and fragility discussed in this thread.
Until that exists, we'll keep building increasingly clever hacks on top of a 30-year-old document format that was never designed for machine consumption.
> The web already went through this evolution once: we went from screen-scraping HTML to structured APIs. Now we're regressing back to scraping because agents need to interact with sites that only have human interfaces.
To me, sites that "only have human interfaces" are more likely that not be that way totally on purpose, attempting to maximize human retention/engagement and are more likely to require strict anti-bot measures like Proof-of-Work to be usable at all.
Great news to all of us keenly aware of MCP's wild token costs. ;)
The CLI hasn't been announced yet (sorry guys!), but it is shipping in the latest v0.20.0 release. (Disclaimer: I used to work on the DevTools team. And I still do, too)
Been using MCP tooling heavily for a few months and browser debugging integration is one of those things that sounds gimmicky until you actually try it. The real question is whether it handles flaky async state reliably or just hallucinates what it thinks the DOM looks like?
Very cool. I do something like this but with Playwright. It used to be a real token hog though, and got expensive fast. So much so that I built a wrapper to dump results to disk first then let the agent query instead. https://uisnap.dev/
Will check this out to see if they’ve solved the token burn problem.
my workaround for this was to make a wrapper mcp server which uses claude haiku to summarize the page snapshot returned in the response of each playwright mcp call, and that has worked pretty well for me: https://github.com/jsdf/playwright-slim-mcp
Been using this one for a while, mostly with codex on opencode. It's more reliable and token efficient than other devtools protocol MCPs i've tried.
Favourite unexpected use case for me was telling gemini to use it as a SVG editing repl, where it was able to produce some fantastic looking custom icons for me after 3-4 generate/refresh/screenshot iterations.
Also works very nicely with electron apps, both reverse engineering and extending.
One tip for the illegal scrapers or automators out there. Casperjs and phanthomjs are still working very well for anti bot detection. These are very old libs no longer maintained. But I can even scrape and authenticate at my banks.
I've been using the DevTools MCP for months now, but it's extremely token heavy. Is there an alternative that provides the same amount of detail when it comes to reading back network requests?
It's probably not fully optimized and could be compacted more with just some effort, and further with clever techniques, but browser state/session data will always use up a ton of tokens because it's a ton of data. There's not really a way around that. AI's have a surprising "intuition" about problems that often help them guess at solutions based on insufficient information (and they guess correctly more often than I expect they should). But when their intuition isn't enough and you need to feed them the real logs/data...it's always gonna use a bunch of tokens.
This is one place where human intuition helps a ton today. If you can find the most relevant snippets and give the AI just the right context, it does a much better job.
i'm experimenting with a different approach (no CDP/ARIA trees, just Chrome extension messaging that returns a numbered list of interactive elements).
Way lighter on tokens and undetectable but still very experimental : https://github.com/DimitriBouriez/navagent-mcp
Unfortunately there are like a billion competitors to this right now (including Playwright MCP, Playwright CLI, the new baked-in Playwright feature in Codex /experimental, Claude Code for Chrome...) and I can never quite decide if or when I should try to switch. I'm still just using the ordinary Playwright MCP server in both Codex and Claude Code, for the time being.
I don’t do any serious web development and haven’t for 25 years aside from recently vibe coding internal web admin portals for back end cloud + app dev projects. But I did recently have to implement a web crawler for a customer’s site for a RAG project using Chromium + Playwrite in a Docker container deployed to Lambda.
I ran the Docker container locally for testing. Could a web developer test using Claude + Chromium in a Docker container without using their real Chrome instance?
Interesting. MCP APIs can be useful for humans too.
Chrome's dev tools already had an API [1], but perhaps the new MCP one is more user friendly, as one main requirement of MCP APIs is to be understood and used correctly by current gen AI agents.
This is the exact problem that pushed me to build a security proxy for MCP tool calls. The permission model in most MCP setups is basically binary, either the agent can use the tool or it can't. There's nothing watching what it does with that access once its granted.
The approach I landed on was a deterministic enforcement pipeline that sits between the agent and the MCP server, so every tool call gets checked for things like SSRF (DNS resolve + private IP blocking), credential leakage in outbound params, and path traversal, before the call hits the real server. No LLM in that path, just pattern matching and policy rules, so it adds single-digit ms overhead.
The DevTools case is interesting because the attack surface is the page content itself. A crafted page could inject tool calls via prompt injection. Having the proxy there means even if the agent gets tricked, the exfiltration attempt gets caught at the egress layer.
Yes, I know it likely breaks everybody's terms of service but at the same time I'm not loading gigabytes of ads, images, markup, to accomplish things.
If anyone is interested I can take some time and publish it this week.
I also use it to automatically retrieve page responsiveness behavior in complex web apps. It uses playwright to adjust the width and monitor entire trees for exact changes which it writes structured data that includes the complete cascade of styles relevant with screenshots to support the snapshots.
There are tools you can buy that let you do this kind of inspection manually, but they are designed for humans.
---
My first reaction to seeing this FP was why are people still releasing MCPs? So far I've managed to completely avoid that hype loop and went straight to building custom CLIs even before skills were a thing.
I think people are still not realizing the power and efficiency of direct access to things you want and skills to guide the AI in using the access effectively.
Maybe I'm missing something in this particular use case?
My excuse for not keeping up is that I'm in so deep that Claude Code can predict the stock market.
I'll still publish mine and see if has any value but agent browser looks very complete.
Thank you for sharing!
https://news.ycombinator.com/item?id=47207790
Never tried that level of autonomy though. How long is your iteration cycle?
If I had to guess, mine was maybe 10-20 minutes over a few prompts.
Here is what actually fixed it: https://github.com/yt-dlp/ejs/pull/53/changes
yt-dlp is relatively stable, but still occasionally breaks for long periods. I get the sense YouTube is becoming increasingly adversarial to yt-dlp as well.
I don't know the details, but it doesn't seem like yt-dlp is running the entire YouTube JS+DOM environment. Something like a real headless browser seems like it would break less often, but be much heavier weight. And Youtube might have all sorts of other mitigations against this approach.
I don't know anything about yt-dlp.
It would probably help people who want to go to a concert and have a chance to beat the scalpers cornering the market on an event in 30 seconds hitting the marketplace services with 20,000 requests.
I can try to see if can bypass yt-dlp. But that is always a cat and mouse game.
I use Patchright + Ghostery and I have a cleaver tool that uses web sockets to pass 1 second interval screenshots to the a dashboard and pointer / keyboard events to the server which allow interacting with websites so that a user can create authentication that is stored in the chrome user profile with all the cookies, history, local storage, ect.. in the cloud on a server.
Can you list some websites that don't require subscription that you would like to me to test against? I used this for Robinhood and I think Linked in would be a good example for people to use.
As someone that does heavy agentic coding (using basically all the tools), this is so far from the truth. People claiming this have probably never worked in large enterprise environments where things like authentication, RBAC, rate limiting, abuse detection, centralized management/updates/ops, etc. are a huge part of the development and deployment workflow.
In these situations you can't just use skills and cli tools without a gigantic amount of retooling and increased operational and security complexity. MCP is really useful here, and allows centralized eng and ops teams to manage their services in a way that aligns with the organizations overall posture, policies, and infrastructure.
> Google is so far behind agentic cli coding. Gemini CLI is awful.
This part I totally agree. It's really hard to express how bad it is (and it's really disappointing.)
Some people will push back on this. They are holding out hope that the recent improvements Anthropic has made in this regard have improved the context rot problem with MCP. Anthropic's changes improve things a little. But it is akin to putting lipstick on a pig. It helps, but not much.
The reason MCP is dying/dead is because MCP servers, once configured, bloat up context even when they are not being used. Why would anybody want that?
Use agent skills. And say goodbye to MCP. We need to move on from MCP.
When folks say MCP is dead, I don't get it. What other alternatives exist in place of MCP? Arbitrary code via curl/sdks to call a remote endpoint?
MCP's remaining moats I think are:
- No-install product integrations (just paste in mcp config into app)
- Non-developer end users / no shell needed (no terminal)
- Multi-tenant auth (many users, dynamic OAuth)
- Security sandboxing (restrict what agents can do), credential sandboxing (agents never see secrets)
- Compliance/audit (structured logs, schema enforcement)?
If you're a developer building for developers though, CLI seems to be a clear winner right
But since when has this industry done the right thing informed by wisdom and hindsight?
It is hard to say nowadays, when things change so quickly
What about all the CLI tools not baked into the model's priors?
Every time someone says "extensibility mechanism X is dead!", I think "Well, I guess that guy isn't doing anything that needs to extend the statistical average of 2010s-era Reddit"
https://github.com/pasky/chrome-cdp-skill
For example, I use codex to manage a local music library, and it was able to use the skill to open a YT Music tab in my browser, search for each album, and get the URL to pass to yt-dlp.
Do note that it only works for Chrome browsers rn, so you have to edit the script to point to a different Chromium browser's binary (e.g. I use Helium) but it's simple enough
edit: upon rereading, I now realize the (different) prompt injection risk you were calling out re: the handoff to yt-dlp. Separate profiles won't save you from that, though there are other approaches.
https://news.ycombinator.com/item?id=47207790
But if I understood the original commenter's use case, they're just searching YT Music to get the URL to a given song. This appears[0] to work fine without being logged in. So you could parameterize or wrap the call to yt-dlp and only have your cookie jar usable there.
[0]: https://music.youtube.com/search?q=sandstorm
[1]: https://music.youtube.com/watch?v=XjvkxXblpz8
Chrome's 'allow pasting' gets ignored reflexively by most users anyway. If this agent can touch DevTools the attack surface expands far faster than most people realize or will ever audit.
I think using the Pi coding agent really got me used to this way of thinking: https://mariozechner.at/posts/2025-11-30-pi-coding-agent/#to...
Also. AAarrgh, my new thing to be annoyed at is AI drivel written slop.
"No browser automation framework, no separate browser instance, no re-login."
Oh really, nice. No separate computer either? No separate power station, no house, no star wars? No something else we didn't ask for? Just one a toggle and you go? Whoaaaaaa.
Edit: lol even the skill itself is vibe coded:
Lightweight Chrome DevTools Protocol CLI. Connects directly via WebSocket — no Puppeteer, works with 100+ tabs, instant connection.
I feel like there's nothing fucking left on the internet anymore that is not some mean of whatever the LLM is trained to talk like now.
HN is becoming close to unusable, and this isn’t like the previous times where people say it’s like reddit or something. It is inundated with bot spam, it just happens the bot spam is sufficiently engaging and well-written that it is really hard to address.
DevTools MCP and its new CLI are maintained by the team behind Chrome DevTools & Puppeteer and it certainly has a more comprehensive feature set. I'd expect it to be more reliable, but.. hey open source competition breeds innovation and I love that. :)
(I used to work on the DevTools team. And I still do, too)
What we actually need is a standard for websites to expose a machine-readable interaction layer alongside the human one. Something like robots.txt but for agent capabilities — declaring available actions, their parameters, authentication requirements, and response schemas. Not scraping the DOM and hoping the AI figures out which button to click.
The web already went through this evolution once: we went from screen-scraping HTML to structured APIs. Now we're regressing back to scraping because agents need to interact with sites that only have human interfaces. A lightweight standard — call it agents.json or whatever — where sites declare "here are the actions you can take, here are the endpoints, here's the auth flow" would eliminate 90% of the token waste, security concerns, and fragility discussed in this thread.
Until that exists, we'll keep building increasingly clever hacks on top of a 30-year-old document format that was never designed for machine consumption.
Citation needed.
> The web already went through this evolution once: we went from screen-scraping HTML to structured APIs. Now we're regressing back to scraping because agents need to interact with sites that only have human interfaces.
To me, sites that "only have human interfaces" are more likely that not be that way totally on purpose, attempting to maximize human retention/engagement and are more likely to require strict anti-bot measures like Proof-of-Work to be usable at all.
Which is called ARIA and has been a thing forever.
Great news to all of us keenly aware of MCP's wild token costs. ;)
The CLI hasn't been announced yet (sorry guys!), but it is shipping in the latest v0.20.0 release. (Disclaimer: I used to work on the DevTools team. And I still do, too)
This is incorrect. Plenty of people have run the numbers. Tool search does not fix all problems with MCP.
https://news.ycombinator.com/item?id=47207790
https://github.com/microsoft/playwright-cli
Odd that this article from Dec 2025 has been posted to the top of HN though
Will check this out to see if they’ve solved the token burn problem.
Favourite unexpected use case for me was telling gemini to use it as a SVG editing repl, where it was able to produce some fantastic looking custom icons for me after 3-4 generate/refresh/screenshot iterations.
Also works very nicely with electron apps, both reverse engineering and extending.
This is one place where human intuition helps a ton today. If you can find the most relevant snippets and give the AI just the right context, it does a much better job.
The problem with MCP is that you are paying the price in token usage, even if you are not using the MCP server. Why would anybody want that?
And no, the tool search function recently introduced by Anthropic does not completely solve this problem.
I ran the Docker container locally for testing. Could a web developer test using Claude + Chromium in a Docker container without using their real Chrome instance?
Chrome's dev tools already had an API [1], but perhaps the new MCP one is more user friendly, as one main requirement of MCP APIs is to be understood and used correctly by current gen AI agents.
[1]: https://chromedevtools.github.io/devtools-protocol/
Hoping from some good stories from open claw users that permanently run debug sessions.
The approach I landed on was a deterministic enforcement pipeline that sits between the agent and the MCP server, so every tool call gets checked for things like SSRF (DNS resolve + private IP blocking), credential leakage in outbound params, and path traversal, before the call hits the real server. No LLM in that path, just pattern matching and policy rules, so it adds single-digit ms overhead.
The DevTools case is interesting because the attack surface is the page content itself. A crafted page could inject tool calls via prompt injection. Having the proxy there means even if the agent gets tricked, the exfiltration attempt gets caught at the egress layer.