· 4 min read · ai
Web exploring tools for my AI agents
WebSearch, Playwright MCP, and Defuddle comparison
The three tools I used
I had a lot of fun this weekend experimenting with AI on some personal projects. It made me think about the different ways my agents can explore the web. I looked a little bit into it, took a few notes and here we are.
Three tools, three very different approache. Each with its own use cases, strengths and limits.
- WebSearch gives your agent a way to search the web. There are MCP servers for it (it’s also built into many popular AI coding agents).
- Playwright MCP is an MCP server made by Microsoft. It gives your agent control of a real browser.
- Defuddle CLI is a CLI tool made by Steph Ango (Obsidian’s CEO). It extracts text from web pages. The agent uses it through a skill (given by the creator himself).1
Links: Microsoft Playwright MCP ↗ - Defuddle CLI ↗ - Defuddle Skill ↗
Comparison
| WebSearch | Playwright | Defuddle | |
|---|---|---|---|
| Primary Purpose | Search the web | Control a browser like a real person | Get clean text from web pages |
| Output Format | Short AI summaries with links | Screenshots, page code, logs | Clean text without navigation or ads |
| Tokens Added to Context | Low (summary + links) | Medium to High (snapshot to full page code) | Low (cleaned text only) |
| Speed | Fast (API call + AI summary) | Slow (page load + rendering) | Fast (HTTP request + text cleaning) |
| JavaScript Support | No | Yes | No |
| Login Support | No | Yes | No |
| Limitations | Can’t interact with pages or logged-in content | Slow to start, uses lots of tokens, needs browser | Can’t interact with pages or logged-in content, some sites block it (403) |
| Best Use Cases | Discover resources | Test websites, fill forms, screenshots, logged-in sites | Read docs and articles |
More Considerations
WebSearch (MCP) uses AI to read multiple sources and write summaries for you. You never get the original content in your context, only the AI’s summary and links. The meaning can change, and you don’t really know where each idea comes from. Both the search and the summary are processed by AI before returning it to your agent. So that’s an extra AI step compared to just fetching a page, and one I’d like to avoid when possible.
Playwright (MCP) shows you the real content exactly as the author wrote it. It starts a full browser and loads all page resources, which obviously is on the heavier side. Then the full page code, screenshots, and logs get added to your AI context, using many tokens and costing more to process.
Defuddle (CLI) outputs to you the real content exactly as the author wrote it too. It makes one simple request to get the page and clean the text (remove ads and navigation), which uses the least energy of all three tools to fetch the content. Only the cleaned text gets added to your AI context, and so using very few tokens.
Final words
I’m using them to improve my research workflow with AI. They each do different things as we saw above and they works well together: WebSearch finds resources, Defuddle reads them, and Playwright steps in when the page needs JavaScript or a login.