Sherin (agentberlin.ai)

I am talking about any client that’s using MCP. The protocol doesn’t have a mechanism that allows writing to a file. You could prompt Claude code to write a program to call the same MCP endpoint and then write to a file but then you’ll have to instruct on handling oauth etc. it may be just works, I haven’t tried it but fundamentally MCPs aren’t built for this purpose (yet)

@__boatbuilder__ @suganthan @keywordinsights I believe you are talking about Claude Desktop or other LLMs. Claude Code can leverage the "tooling" they have available and directly store the output to a file to exactly prevent this. Then the model uses grep (as another tool) to understand the file.

MCPs are so bad for SEO. Why would anyone use MCPs for any SEO tasks? We used a couple of them, we built like 6 of them. But a single keyword universe typically is like 10s of thousands of keywords and similarly, all the SEO data. We spend way too much time avoiding the hallucination that we said enough is enough - no more MCPs

Yeah, I am not sure if you understand what happens under the hood when you ask an LLM that. It reads the response from MCP tool and then use the write_file tool to which the LLM has to pass the copy of the data it read in that format itself. As far as I know, MCP doesn’t have a spec that allows LLM to ask it to directly write to a file.

@__boatbuilder__ @suganthan @keywordinsights using Claude Code, you can ask it to directly save the output of the MCP call to a CSV. Then, Claude Code uses grep to get only the info it needs, plus understanding the data structure getting the first lines of the CSV :)

@guifreballester @suganthan @keywordinsights Yeah, but LLM writes the csv based on the data it has in it’s context that it got from MCP right? Or did you find a way to bypass the data going to LLM context at all and do direct MCP -> local file

@suganthan @__boatbuilder__ @keywordinsights I'd suggest to keep using MCP (only way to have reliable data without hallucinations). The magic here is to ask LLM to save intermediate results to CSV or MD files. This way you save context. Also, the SEO MCP revolution is not in Claude Desktop but in Claude Code!

@suganthan @keywordinsights Sweet! so skill instruct the LLM to talk to MCP in a specific way?

@__boatbuilder__ Fully agree. When i loaded Ahrefs and DataforSEO into Claude it crashed completely. I couldn't even get a message to work due to hitting context window. We built the MCP for @keywordinsights but switched to a skill now. It's much easier for non tech folks to use the tool 😃

@suganthan Brilliant idea. I mean, everyone who had using LLM in prod has some version of managing context based on situation. “Skill” is a standardization of that! That’s my view. What do you think of MCP and skills?

@__boatbuilder__ Thoughts on skills?

Claude Code has 5x'd my productivity. But I'm not spending less time with code. I'm spending more. AI writes most of it now, which means I'm planning more, reviewing more, iterating faster, thinking deeper and shipping quicker than ever.

Analytics is broken and most teams aren’t ready for what’s coming next. Your website now gets three types of traffic: humans, indexing bots, and AI agents visiting on behalf of users. We track humans well. We mostly ignore indexing bots. And we barely track AI agents at all. Most AI agents don’t execute JavaScript, which means your analytics never fire. Your site can be read, cited, and recommended by AI systems without you ever seeing a pageview. You can’t optimize what you can’t measure and right now, a growing part of your audience is invisible.

Why hasn’t AI transformed marketing the way it transformed coding? Because AI works best where outcomes are verifiable. Code is verifiable, you run tests and check output. Marketing isn’t. You publish content, optimize pages, then wait weeks or months without knowing what actually worked. When results change, you can’t tell if it was content quality, links, timing, or an algorithm update. The signal is buried in noise. That’s why most “AI marketing” today is just faster manual work. Without clear feedback loops, AI can’t really learn, it can only guess faster.

99% of founders are building companies wrong. This is why, after 3 years of development and 14 months securing approval from the European Commission for Educational Content Compliance... I'm pleased to announce my MasterClass: "The Art of Compliance" In this class I will teach you: - Why the cookie consent banner isn't blocking your product—it IS your entire product - Why I rejected Stripe at pre-seed for having ZERO certifications (the Collison brothers personally called me, I hung up) - Why I wake at 3am to physically touch my SOC 2 report before returning to sleep - Why you must hire a Chief Compliance Officer on DAY ONE—before your first engineer, before your first line of code Stay tuned. Access details pending approval from 11 European data protection authorities.

(1/2) Today we’re starting to test a new way to seamlessly go deeper in AI Mode directly from the Search results page on mobile, globally. This brings us closer to our vision for Search: just ask whatever’s on your mind – no matter how long or complex – and find exactly what you need. You shouldn’t have to think about where or how to ask your question.

@pmitu Didn't know Github copilot got better to be included in this list. Claude Code any day. I am constantly trying all of them for different things. Nothing beats claude code's capability and the UX and since it has both terminal and web interface, convenience is just unbeatable

Hey, vibe coders, what's your choice: - GitHub Copilot - Claude Code - ChatGPT - Windsurf - Lovable - Cursor - Bolt - v0 - ? 👨‍💻 Choose your fighter

@SimonHoiberg Surprised that nobody mentioned outerbase.com - Runs on browser - Now under @Cloudflare

I use pgadmin everyday and I hate it so much. The UI/UX is is making me sick 🤢 I want Supabase-grade UI, but I DON'T want all the underlying tooling. Literally just the UI so I can inspect my DBs and do queries. And I can't for the life of me find this tool. This doesn't exist? Who's gonna create it? I'd pay for this on a lifetime deal.

@lilyraynyc How do you keep finding interesting things!

In case anyone needs this

Kimi-k2-thinking is not a joke. Switched our Reddit research agents from GPT-4/Claude to Kimi K2. We're running 50+ agent loops doing complex multi-step analysis - digging through Reddit for SEO insights, user pain points, untapped communities. Not simple queries, actual deep research work. Cost dropped 75%. Same results.

@DeepakNesss Great. I need to get a certificate for the desktop app so it doesn’t show security warning to you.

@__boatbuilder__ macOS

My ScreamingFrog license expired 2 days ago and I don't want to pay $279 just to use it a few times a year. So... what are the cheaper/free alternatives?

@marclou Give Claude code a try? It executes commands, write test files, check the output, fix the issues, then delete the test files - the whole developer package

Claude sonnet 4.5 debugs with console log instead of brute forcing random guesses like Claude 4. Like it was meant to be.

@lewiscarhart Man! Insane respect

we're close to hitting $1M ARR after just a few months every single day, people ask me: "how is it possible" here's how:

OpenAI hiring a content strategist for half a million dollars is a clear signal to founders: don’t quit on SEO.