I have tried a bunch, but LLMs seem to be way too unstable sometimes. For example I explicitly say what MCP to use and in the end it does not. It rarely calls MCP tools on its own.
Context7 is great for combating hallucinations and helps guide the agent within the environment. I add it to every project, but rarely use it. Agent never uses it on its own despite having it in its rules.
DebugMCP is great for debugging but not perfect. Barely use it.
Dart/Flutter has a good MCP server.
Laravel has a decent one too.
I dont vibecode much, I use AI as a research tool and these are the only ones I found to be useful for my interactions.
I used them quite religiously at the start (Playwright, Azure) but have gone off them because:
Theres an Inherent black box risk of using them especially for larger companies (even if they are “read only” I am still a bit uncomfortable risking it)
When I’m comfortable enough to use them, agents are smart enough to just work out the appropriate CLI commands which are good enough.
Playwright/Google DevOps MCP seems to be a bit slower than it used to be. Like many things in this game I can’t definitively prove it, but it’s something I’ve noticed.
Often it’s quicker just to do things yourself, it’s tempting to just delegate everything to an agent but we all get distracted on X/YouTube/life/whatever and I forget about the task that’s running. Not specific to MCPs though !
Final thing I did find the Azure MCP very useful at times for configuring pipelines and app services, in particular containerised ones where I don’t have much experience. Back then I didn’t use terraform
I use just one server, which has over a dozen tools - it uses context-compression (aka progressive-discovery) so it’s safe to leave all tool on at all times, and it doesn’t waste context.
It can do EVERYTHING with that - crazy in-browser work, full-stack devops, IoT, complete desktop automation, CAD, 3d-printing, …
You get the BEST results, when you give it tools it can use to test what it’s doing (like ssh/desktop-automation, good browser…) - and use Opus 4.5.
Thank you very much for your time and for sharing your experience — the insights you mentioned are genuinely valuable, especially your point about LLMs not reliably using MCPs even when explicitly instructed.
If you don’t mind, could you please share the names or links to the Flutter and Laravel MCPs you referred to?
Also, if there are any other MCP tools you’d recommend or find useful in similar contexts (even beyond direct coding), I’d be very grateful to hear about them.
I truly appreciate these valuable insights and your practical perspective.
Thanks again for taking the time to share
Sorry, my mcp experience was very bad. I’ve tried github, atlassian, dbhub, figma, playwright and here are my takeaways:
github - half of the time it would spend token on recognizing the user and checking permissions, when it was not needed at all, prefer using github cli gh over this mcp
atlassian - worked for most of the part, but the super annoying stuff of authenticating daily, sometimes even multiple times a day is a nightmare, prefer copy pasting ticket and confluence stuff instead
dbhub - how many times it never used the db even though I used to ask the models to (Gemini 2.5 pro times), it used to fail, then I had to tell it to retry, then it would work, but get stuck before giving me the output, enough of it, I manually look at dbs using Tableplus now
figma - it worked for some time, but i never found it helpful much - as it could not grasp selections inside figma designs only the whole design. Models are become better at making UI by just attaching the screenshot of the figma ui. the minute details I add manually.
playwright - i used it to test my ui, it works well - but it won’t use my chrome instance and create a new one, for the new one i have to setup login and cookies and stuff which is annoying. I use cursor’s browser use instead, its still the same issue, but its all inside the ide with good selection tools
To summarise, prefer raw methods instead like cli, direct access - mcps are just an abstraction of those which waste our times.
If anyone is using these MCPs to the fullest - do let me know.
this looks interesting. might want to elaborate the upsides in combination with cursor ? any workflow examples ? almost all mcp’s shown on your site are not really of interest for me, but maybe im getting you wrong ?
Have a look at Playwriter.dev on GitHub which allows you to use an existing tab rather than a test profile and is mentioned around on X. It has some other nice features too.
Like you I’ve gone off it a bit though albeit my roles changed quite a bit recently.
What are you building/doing with Cursor? Those tools cover every aspect of everything you could want, packaged with progressive-disclosure so you can leave them running all the time (saves mass token-usage - one of your goals). The biggest game-changer is testing and devops: you can “one shot” almost anything, walk away, and come back to it finished and live. Even crazy stuff, it can handle in one-prompt.
I use Honeybadger MCP server. It was unreliable at first but the HB team fixed it and it’s pretty great now, I can post in a link to a HB fault and then Cursor can pull all the details for it (and usually manages to fix it without any extra guidance).
I usually have to specifically tell the agent to use the HB MCP server though or it won’t figure that out on its own… and some LLMs are better than others. 5.1 Codex generally nails it first try.
,