Marco Pfeiffer

Getting déjà vu with Gemini 3.0. 2.5 Pro dominated benchmarks too, but struggled with edge cases in practice. Now 3.0 tops the charts again… let’s see if it actually handles real-world scenarios better this time.

Interesting find: Perplexity Desktop has MCP support 👀

Tested it with my ClickUp server - works for basic queries, but no complex workflows. Makes sense for a search tool though. What would really fit: syncing data sources so it searches my tasks alongside web results.

Tried Antigravity's planning mode:

First attempt (Gemini 3): unknown error
Second attempt (Claude 4.5): what planning mode? *applies changes instantly*

Not a great start.

My TYPO3 MCP server now has proper setup docs for n8n and manus. n8n is perfect for visual workflow automation, while manus can handle complex multi-step tasks much better than other clients - ideal for bulk content operations.

Gave Manus AI a shot with my TYPO3 MCP server. Had it create 40-50 news entries - structured everything, wrote docs for missing pieces, generated Python scripts. Painfully slow, but impressive how well it tackles complex workflows.

Just realized Perplexity has had (local) MCP support for nearly 4 months. Somehow completely missed this despite using it regularly. Time to actually explore what you can do with it.

I completed a complex task with Opus and it performed well. However, almost half of my weekly quota of Opus has been used up. You really have to use that sparingly.

TIL that SQL BETWEEN is order-dependent. First value must be ≤ second.

I use it so rarely that I never noticed 🤷

ClickUp launched official MCP support. My own implementation is now obsolete, but this is great news – official APIs and support will make the ecosystem way more stable for everyone.

Hetzner dropped prices again. Already the best value out there.

Am I crazy for wanting even smaller options? 1c/2GB at ~€1.69/month would be ideal for single-purpose services. Not everything needs 4GB.

Unexpected perk of running macOS beta: it forces me to actually reboot my machine occasionally

Just hooked up Claude Code to our BitBucket CI/CD pipeline.
After each deploy, it analyzes commits and updates our ClickUp changelog automatically.

No more "we'll document this later" (narrator: they never did).

PMs can still edit manually, and Claude keeps the format consistent.

Claude's vision capabilities are seriously underrated. Just drop in a screenshot of a ticket, chat conversation, or even messy URLs and tokens - it reads everything perfectly. I keep forgetting to use this more.

Sonnet 4.5 is genuinely impressive at UI testing. Just watched it dive into MySQL to verify the actual data before reporting back. Goes way beyond surface-level checks.

Now that Anthropic released Sonnet 4.5, I can see some Max20 users downgrading. If it’s really that good, Max5 should be plenty.

Remember when Star Trek's Computer just created interfaces on voice command? We're getting there. Claude Imagine generates full UIs in real-time - fascinating approach.

Claude Desktop describes Sonnet 4 as a Legacy Model.
Spoilers.

Is the ChatGPT Mac app actually maintained? Feels like new features arrive late (if at all). No connectors, no model switch in projects, Custom GPTs just kick you back to the web.

When copying text comes with a progress bar, you know your day has taken a turn.

Rebranding Remote Desktop as Windows App and then leading with “security” is the clearest sign that marketing doesn't know what this product does.

With Claude Pro I get about the same done as with ChatGPT Plus + Codex. The real pain is peak time: slow code, cut-off answers. It's really fun on weekends though.

Had to check how overpriced Anthropic’s Team Plan for Claude Code is.
Answer: ~5×.
If the regular seat had the usage of the Premium seat, I’d happily buy it for the whole team.
Instead, it just feels… absurd.

I downgraded my Claude plan to Pro (so no more Opus in Claude Code). I thought I’d rely more on Codex… but now I’m using Claude Sonnet way more than before. Was it always this good? Was I blinded by Opus?

Zoom's 40-minute limit on free accounts is actually brilliant for time-boxing meetings. Slack huddles fill the gap nicely for quick 1:1s.

Funny thing: ever since I started testing Codex more, Claude Code decided to become twice as fast and reliable again.
Almost like it doesn’t want me to switch. 😅

Cloud Code sneaked a rainbow animation into their terminal UI 🌈
Always wild to see how far you can push the boundaries of a terminal app.

Just figured out why Codex struggles so much with tickets: it can’t really read text in screenshots. Two or three images and the whole GPT-5 context is eaten up, yet it still doesn’t know what the screenshot said. Claude Code seems to handle this better. OpenAI really has to improve Codex here.

I already miss Safari’s compact tabs.
With the traditional layout, two single-tab windows can’t be merged because the tab bar is hidden. That feels like a step back.

Nextcloud’s window radius nearly matches macOS Tahoe’s new style. Nearly. And that ‘nearly’ keeps bothering me.

Everyone probably already knew this, but I didn’t.
Been chasing a Node.js memory leak for a while. Turns out it was this:
undici.nodejs.org#/?id=garbage...
“Always either consume or cancel the response body.”