3 min read

I migrated from Cursor to Claude Code and saved $400

My honest take after a week with Claude Code, OMC, and Serena. The full setup with MCPs, workflows, and what actually works.

My take

It was one of the best things I have done.

Before, I used Cursor, just Cursor and a few MCPs, and I trusted how it managed tokens, compressed context, handled agents, and everything else. But last month I burned $600 in tokens on Cursor, and honestly the code quality was not great, because on Cursor I always used Auto mode, which probably runs their “Composer” model most of the time.

The migration

When I moved to Claude Code, I hit an obvious problem first: Claude Code does not support multi-repo workspaces, at least not yet. That let me down a bit, but I worked around it. Using the Claude Code CLI (which I found much better than the UI, since I run dozens of subagents at once and UIs start to choke) I created a parent folder and dropped my sub-repos inside it. That solved the problem, and as surprising as it sounds, Claude handles it well.

But a couple of things were still missing in Claude: an agent orchestrator and workflows. Workflows are, well, work flows. For example, I always like to ask the AI to lay out a plan for solving the task and only then implement it in parts. That is when I ran into something called Ruflo.

The Ruflo trap

At first glance, it promises everything. But after a few hours of using it, I started to notice things did not work like “in the ad.” Until I cloned the repo and began finding all sorts of gaps: functions wired to nothing, features that were just mocks.

That is when I decided to dig in, and look what I found: Ruflo / Claude-Flow: 300+ MCP Tools Exposed, 99% Theater, 1% Real, 5x Token Waste, and roman-rr had reached the same conclusion I had reached minutes earlier.

At that moment I realized how many people who understand nothing about AI are selling AI to you like a golden hammer that solves all your problems without you knowing anything. If I were not a “real” developer, I would not have noticed.

The setup that actually works

So I decided to go all in and figure out how to actually improve my Claude Code. After a lot of digging, here is what I landed on:

Claude Code CLI: for UI performance when running many agents.

A real agent orchestrator: OH MY CLAUDE CODE: real workflows, commands that work, no theater.

Plugins (MCPs):

  • context7: up-to-date docs for any lib straight into context
  • exa: real web search
  • filesystem: filesystem access
  • memory + qdrant-memory: persistent memory across sessions
  • serena: the cherry on top (semantic code analysis, symbol navigation, safe refactoring)
  • aws-api, aws-docs: for those who work with AWS

Commands: OMC ships several workflow commands like /team, /ralph, /ralplan, /ultrawork, and many more to run specific tasks in specific ways.

The result after 1 week

I paid $200 for the Max 20X plan, used Sonnet 4.6, used Opus 4.7, and I really pushed it:

wk: 64%, I used it way more than I ever used Cursor, and I did not even manage to hit 100% of my weekly limit.

LONG LIVE CLAUDE CODE, and OMC, and SERENA!

Conclusion

If you are paying a lot for Cursor and you are not happy with the quality, it is well worth trying Claude Code with this setup. There is a learning curve, but the payoff matches it.

If you want me to explain in detail what Serena, OMC, qdrant-memory, and memory do, ask in the comments and I will write a dedicated post.

Originally published on r/brdev.

Comments