Chat, CODE IDEs, AI Orchestration Tools, Factories, MCP servers, Sub Agents Pol - What do you use? Discussion.

What tooling do you use to build?

  • Chat

    Votes: 5 100.0%
  • Code IDE

    Votes: 0 0.0%
  • AI Orchestrator

    Votes: 1 20.0%
  • Factory or similar home brew or external

    Votes: 1 20.0%
  • MCP servers

    Votes: 4 80.0%
  • Sub Agents in non factorial configurations

    Votes: 1 20.0%

  • Total voters
    5

secretagentdad

Chief Fed Poster
BuSo Pro
Joined
Jan 29, 2015
Messages
1,311
Likes
1,221
Degree
4
What non AI components are in your stack?

How do you customize your super weapons.

What GPT plans and models are you using?
What is the typical effort level?

This community makes money when it rains.
Its pouring. Sharing is caring. Walk us through your work flows.
We need some good outing to inspire each other.
I'm gonna leave space for some other people to post instead of drowning the thread with a text wall.
 
I am using Claude (ultra, $200/mo) to build a massive site (4,000 pages so far).

I use MCP connections for:
- ahrefs (seo data)
- Airtable (website db lives here)
- Webflow (front end for the airtable data)
- Figma (all my graphics live here)
- Socialbu (LTD that lets you schedule content via MCP - I have 8 accounts set up including a subreddit getting industry news 5x a day)
- Notion (for managing tasks, staying organized)
- SERPAPI for querying google
- Context.dev for scraping

For effort levels it depends on what I am doing. If it's just data extraction, I'm using Sonnet. Opus for most things, and Fable for design, high level planning, and important decisions.

Running all of this out of the standard Claude app on windows/mac. I have lots of scheduled tasks to handle content refreshing, social media posts, and data sourcing.
 
I've tried everything, and anything.

I make time in my day specifically for this task.

The new model you heard on YouTube? The github skill you saw on X? The harness or concept you found on TikTok? I've prob already tested it 3 months ago and moved on. Not saying that to brag, but this is how I absorb learning as a AuADHD.

I've learned, like any other topic in life.. you gotta settle on your flow. Stop chasing shiny new and go deep in the rabbit hole on what you settle into.

I don't want to be locked into anything, that can be taken away from me later... or prices raised 5x over the next year. I learned that from having SaaS subs and other things years and years ago.

As far as models/Ai:
1. Claude $200/m plan - mostly use Opus and Medium thinking. I almost never run out of my 5-hour ( now weekly ) limit. This is my goto general daily stuff.
2. Codex/ChatGPT $20 plan - mostly use this as backup to check Opus outcomes. I dont even know what model or thinking level it's set to.
3. Alibaba sub - Mostly for overflow, but Im finding Im not using it much. Might cancel.
4. GLM coding sub - for GLM 5.2/3 - I use this for coding and starting a new code/app idea. i can get something fleshed out well and then later let Claude or Codex take over to fine tune it.
4. OpenRouter when I want to test a new model someone talks about like deepseek 4.1 flash and others.
5. LMStudio/OMLX/llama.cpp - local models on my 128gb M3 laptop. Have stuck with mostly gpt-oss-20b and 120b, qwen MoE models, and gemma models on this.


Harness:
1. Im mostly using Claude or Codex as is when I use those. Typically as "projects" instead of one off queries. I prob use Chat and Code equally.
2. Hermes - Ive been using Hermes for a long ass time. I mostly use this with the GLM stuff and coding and building out new ideas.
3. I've tried all the new and flashy stuff like the deepseek harness, heder, tmux, cmux, and all the bots like grokbot, dots, hermes bots, etc too.


Data storage/skills:
1. I was all into LLM Wikis, SQL databases ( BM25 FTS ), graphs, vector DBs, grep file searches, NotebookLM, and others as a RAG system. I was building these out and using them separately and combined for results. I've tried it all. I even weight programmed my own local LLM with editing the open weights, adding new weights, layering on LoRA, etc. I found them all to be helpful, but very "not up to my standards" for various reasons even with combined.
2. Ended up moving away from RAG largely because of context rot and token usage, especially on models that dont have 1m token limits. But even on 1m I find it difficult to do the work I do even with sub agents. So I moved to just having very very precise and distilled skills and got rid of RAG mostly and weight programming my own open models for very specific usage. 1 edited model for X, another for Y, etc. I think the biggest win is still the very very in tune skill, but the edited weight models are picking up a lot of stuff I dont have to waste tokens on.


Flow:
1. I tend to keep all my stuff outside of the LLMs and Harness as much as I can. Meaning shared folders ( across my Macs ) and cloud ( Google Drive ) or S3 type budgets ( backblaze, Amazon, etc ) and just have the Ai reach into those as needed. So skills, data, outputs, etc. This prevents vendor lock-in and ensures I can use any model or harness I want in the future.


Going deep now into specifics, instead of chasing shiny:
1. Im pretty much only coding in Go and Sqlite now. Sometimes I might reach for Python.. but I gave up on trying to build in PHP, Rust, Erlang, and all the databases. Going deep in 1 or 2 is the way and Ive decided that being deep in 1 or 2 is better than chasing and having code in lots of others. Go is so simple and short, same with SQLite. How you gonna mess up Go and SQLite? Bad code is bad code always, but Im talking about the ways to achieve something in Go.
2. I decided to stick with Hermes as the harness ( unless Im deep in claude or codex and then I just use theirs ). Anything I need a scheduled job for, cron, etc.. I tend to stay out of Claude or Codex because I don't want to be locked in and then depend on that.
3. I make my own skills and I go very deep in them. When you look at other people's skills, even if they have 10m stars on github, you find they are very generic and very "junior". Im not down with that.
4. Im not using MCPs or any of the connectors in the Harness. I want my own code, my own CLI, my own tooling. Again for no vendor lock in. If I need an external tool, I write my own CLIs to it and connect to it.
5. For stuff outside of coding, its very very deep and precise. Im not using the "general knowledge of the Ai". Im always fine tuning for the topic and work. An example would be marketing, Im not using the general Ai knowledge for it. This is how I handle a lot of topics I deal with.


I tend to use Ai for my own use cases. Not for clients or others. Make my day simplier, easier, more relaxed first. When it does that, and I control everything from A-Z in it with no lock-in, then I open it up for others to use and charge them for it while I ensure Im in total control.


I've used Ai for everything you can think of. From rewriting and programming Linux Mint, writing new code on an old iMac without Metal to use Metal libs, copywriting, projecting and auditing my taxes ( with all the personal and sensitive stuff blacked out and removed ), debating me on topics, creating apps ( web and desktop based ), watching competitors, automating social media, legal contract overviews, creating video and image ads, finding leads for a business, finding gaps in codes and ideas and knowledge, and more.


I def take everything with a grain of salt though. I'm not blindly trusting it 100% like some other people do. I always apply my own due diligence to everything with it.


The 1st goal is to free up MY OWN day and time. 2nd goal is to make things simpler. 3rd goal is to make things more accurate and reliable. 4th goal is to control all of it and make sure I own it. 5th goal is to then turn whatever I just did above into revenue.
 
Last edited:
Back