Skip to content
Get App

How to Connect Hermes Agent to Playwright MCP

Kris Torrington shows how to connect the Hermes AI agent to Playwright MCP inside WSL so it can control a real browser for form filling, site navigation, and live Amazon checks.

Key Takeaways

  • Combining Hermes with Playwright MCP enables real browser interaction beyond what web search can do.
  • Snapshot-based page inspection lets the agent identify exact input elements rather than guessing.
  • Reliable browser agents must handle obstructions like pop-ups, inconsistent layouts, and missing data without fabricating results.

What the video covers

  • Setup: open WSL, create a Node project, and install Playwright MCP and Playwright with NPM to get a headless Chromium browser.
  • Verify the Chrome/Chromium executable path, then add a 'playwright' MCP server to the Hermes config YAML using npx with @playwright/mcp@latest, headless, and browser=chrome arguments.
  • Test 1: Hermes opens example.com via Playwright MCP and reports the page title.
  • Test 2: Hermes adds three tasks to the Playwright TodoMVC demo app, identifying the input field from a page snapshot and verifying the results.
  • Test 3: On Rotten Tomatoes, Hermes dismisses a privacy dialog, opens the top three movies, extracts Tomatometer and audience scores, and avoids fabricating missing data.
  • Test 4: On Amazon, Hermes searches for a PlayStation 5 Pro, extracts the price, and checks buy-box elements to determine whether direct purchase is available or only third-party sellers exist.
  • Key distinction: web search retrieves information, while Playwright MCP adds an action layer letting the agent observe, interact with, and verify live pages.
  • The approach avoids bypassing CAPTCHAs or anti-bot systems; it only inspects what a normal browser session can see.
  • The video closes by previewing a follow-up on hosting Hermes on a VPS for 24/7 operation.

Answers

Questions about this video

What is the purpose of connecting Hermes to Playwright MCP?

It gives Hermes an action layer so it can control a real browser, click buttons, fill forms, and verify results on live websites, which web search alone cannot do.

How is the Playwright MCP server configured in Hermes?

The creator adds a 'playwright' server entry to the Hermes config YAML with the command npx and arguments including @playwright/mcp@latest, headless mode, and browser set to chrome.

Does the video show bypassing CAPTCHAs or anti-bot protections?

No. The creator explicitly states the tests do not involve bypassing CAPTCHAs, defeating anti-bot systems, or accessing protected information; they only inspect what a normal browser session can see.

Full Transcript — Download SRT & Markdown

00:00
Speaker A
most AI agents can search the web but the second that you ask them to click a button fill out a form or use a real website they fall apart in this video I'm connecting Hermes to Playwright MCP so it can control a real browser inside of WSL
00:16
Speaker A
on Windows by the end Hermes Agent will add tasks to a web app work through a messy movie site and check a live Amazon listing like a real user would so let's set it up first open Windows Powershell and type WSL
00:33
Speaker A
that drops you into your Linux environment once you're in create a folder for the project by running make Directory Playwright MCP then move into it with the CD command now create a new node project by running NPM in it
00:49
Speaker A
then install Playwright MCP and playwright with NPM this installs the MCP Server Hermes Agent will use along with the browser automation tools running behind it if it asks for your sudo password enter it so Linux can install the extra dependencies
01:09
Speaker A
once that finishes playwright has a headless chromium browser ready to use before we connect it to Hermes run the following command to ensure that Chrome is installed correctly that should return the path to the browser executable now we need to connect it to Hermes
01:28
Speaker A
so open your Hermes config file by running Nano Hermes config dot yaml scroll to the bottom and add a new MCP server called playwright set the command to MPX then add these arguments at play right MCP latest headless and browser equals Chrome
01:50
Speaker A
save the file with control and 0 then press enter and then now finally exit by pressing control and X at this point Hermes can launch playwright whenever it needs to interact with a browser and this is the big difference
02:05
Speaker A
so let's test it start the agent by typing Hermes Chat for the first test tell Hermes use Playwright MCP to open example.com and tell me the page title Hermes sends the instruction to playwright MCP playwright controls the browser
02:25
Speaker A
the browser opens the website and Hermes reads the result back to you now let's give it a task that web search cannot complete for us for this test tell Hermes to go to playwright to do MVC demo site
02:41
Speaker A
and add these 3 tasks record video edit thumbnail and upload tutorial this time Hermes has to do more than read a page it needs to understand the layout find the correct input field enter the task submit them and confirm that the action has worked
02:59
Speaker A
first it navigates to the site then it takes a browser snapshot and scans the page structure instead of guessing where the input box is it identifies the exact reference for the task text box then it types record video
03:15
Speaker A
edit thumbnail and upload tutorial submitting each task as it goes finally it checks that all three items appear in the list and confirms the result directly in chat that is the point where the difference becomes obvious search can tell Hermes what to do MVC is
03:35
Speaker A
Playwright MCP lets Hermes add the task itself Hermes is no longer just reading websites it's observing the page interacting with the interface and verifying that it has completed the task correctly now let's move to a real world website with pop UPS
03:52
Speaker A
changing layouts and incomplete data tell Hermes go to Rotten Tomatoes find the top movies opening this week open the top three movies extract their audience and Tomato Meter scores and give me a quick summary this is where browser agents either prove themselves
04:11
Speaker A
or fall apart the site loads and there may be a privacy dialogue blocking the page a basic scraper might fail immediately but Hermes can recognize that the page is obstructed locate the dismiss button click it and then continue
04:27
Speaker A
next it scans the page to identify the top movie titles then opens each movie one by one and looks for the Tomato Meter score the audience score and any missing or unavailable data the important part is that if a score is missing
04:43
Speaker A
Hermes should not make that one up it checks the page structure confirms what's available and reports the missing information clearly the result is not just a copied list of numbers it is a summary based on what Hermes can actually see on the live page
04:59
Speaker A
now for the hardest part test Amazon Amazon is difficult because the layout is dynamic product listings change constantly and purchase options can vary by seller region or stock status for this test we're not trying to bypass CAPTCHA's defeat anti bot systems or access protected info
05:21
Speaker A
we are simply asking Hermes to inspect what a normal browser session can see so give Hermes this prompt Search Amazon for a PlayStation 5 Pro console find the most relevant listing extract the displayed price and tell me whether the product is available to buy
05:39
Speaker A
directly from Amazon or only through third party sellers Hermes starts by locating the search box and submitting the search it then identifies the relevant product listings checks the title examines the display price and opens the product page but the price alone is not enough
05:59
Speaker A
a product can have a listed price while still being unavailable for direct purchases so Hermes checks the buying area is there an add to cart button is there a buy now button does the page say the item is unavailable
06:14
Speaker A
are the only purchase options come from third party sellers if Hermes finds an odd price a missing button or unclear buy box it does not stop at the first tool error it checks the surrounding layout and visible page text
06:29
Speaker A
to understand what the page is actually telling it then it can give a practical result such as the PlayStation 5 Pro listing is currently not available for direct purchase from Amazon the visible purchase options are from third party
06:43
Speaker A
sellers and that is the real power of combining Hermes with playwright MCP the agent is not just executing a fixed browser script it is observing what is on the page adapting when the page changes checking its own work and reporting what it's found
07:00
Speaker A
in this video we connected Hermes to playwright MCP inside of WSL on a Windows machine then tested four increasingly difficult browser workflows First Hermes read a page title then it completed a task in a live web app then it handled a privacy pop up
07:18
Speaker A
inconsistent layouts and missing data on Rotten Tomatoes finally it inspected a dynamic Amazon listing and reasoned about actual purchase availability the big takeaway is simple Hermes Web search helps it find and understand information player right MCP gives it an action layer
07:37
Speaker A
together they let Hermes research the web navigate live pages interact with browser based tools and verify outcomes in real time so if you enjoyed this video make sure to hit subscribe for more AI related tutorials and if you wanna learn how to host a maze
07:54
Speaker A
on your own VPS check out my next video where I'll show you how to host your Hermes agent on a VPS using hosting a so you can keep running 24 7 outside your local machine I'll see you in the next one
Topics:Hermes AgentPlaywright MCPbrowser automationWSLAI agentsMCP serverweb scrapingRotten TomatoesAmazonNode.js

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

Or transcribe another YouTube video here →