I’ve been using AI coding assistants for a while now. Not because I can’t write code without them — I can. But because they save me time. And time is the one thing I never have enough of.
I started with Cursor. Then I tried Trae. Then Claude Code showed up and everyone lost their minds. And now I’m on OpenCode.
Here’s what actually happened.
The Cursor Era
Cursor was my first real AI coding tool. I got into it because the tab completion was noticeably better than GitHub Copilot. Like, actually useful. Copilot would guess what I wanted and get it wrong half the time. Cursor got it right more often. It felt like it understood what I was trying to do.
Then my company gave us a $500 monthly budget for it. And I went all in. Opus model, unlimited usage, the whole thing.
For a while, it was great. I’d describe what I wanted in plain English and Cursor would generate working code. Not perfect code, but close enough. I could fix the edges and move on.
But here’s the problem — that $500 budget didn’t last. Opus burns through tokens like crazy. I could watch the balance drop in real time. My team started tightening the limits. First we had to request approval. Then we had to justify every expensive query. Eventually, the budget got cut.
I could see it coming. I’m honestly surprised it lasted as long as it did.
The Trae Detour
I tried Trae early on. This was right when it came out, and people were comparing it to Cursor.
Honestly? It wasn’t even close. The gap was obvious from day one. It felt slower, less accurate, and the UX was just… off. I gave it a few days and dropped it.
I don’t have much else to say about Trae. I’m sure it’s gotten better since then. But I never went back.
Claude Code: The Hypetrain I Hopped Off
When Anthropic dropped Claude Code, the entire internet lost its mind. Every tech influencer was posting about it. It was everywhere.
I tried it. Of course I tried it.
Here’s the catch though — I couldn’t use Anthropic’s models directly. Some policy thing. So I went through CC Router, which lets you route Claude Code through Chinese models instead.
It worked. Technically.
But it always felt like I was one step removed from the real experience. Something was off. The responses were slightly slower. The context window didn’t feel as big. It was like using a car that had been imported and converted to right-hand drive — functional, but not quite right.
Also, Claude Code is closed-source. The ecosystem is locked to Anthropic’s models. If you want to use something else, tough luck. That bothered me more than I expected.
I think multi-model agent teams are the future. I don’t want to be locked into one provider. So even though the community kept hyping it up, I walked away.
OpenCode: Where I Landed
I switched to OpenCode and haven’t looked back.
The biggest difference? OpenCode was built for all models from day one. Claude Code only works with Anthropic. OpenCode works with everything. GLM, DeepSeek, Qwen — you name it.
And with Chinese models getting genuinely good, that’s a huge advantage. I’m not kidding when I say DeepSeek-V4 holds its own against Opus. Sometimes it’s even faster. And it costs a fraction of what Opus does.
Here’s my current setup:
Orchestration: OpenCode-Go (GLM 5.2 + DeepSeek-V4 Pro for execution, DeepSeek-V4 Flash for the heavy lifting)
Multimodal: OpenCode-Go with Kimi or Mimo
Plugins: oh-my-opencode-slim + magic-context
Monthly cost: $10 for the subscription + a little API usage
Wait, I should mention — I’m actually paying $5 for the subscription because I keep renewing the trial thing. Not sure if that’s still available. But even at $10, it’s fine.
The plugins make a real difference. Oh-my-opencode-slim keeps the agent behavior under control. The full version is too vibe-coded for me — burns tokens and feels unpredictable. The slim version is tighter. Magic-context handles context management so I don’t have to think about it.
My company still has Cursor credits available. I barely use them anymore.
One More Thing: Hermes
I’ve been using Hermes alongside OpenCode recently. They do different things.
OpenCode is for coding tasks. Writing functions, refactoring, debugging, that kind of stuff. Hermes is for collaboration. Messaging, scheduling, information management — the boring but necessary stuff.
Here’s what I use Hermes for every day:
Daily summaries. I have it scan my IM messages and generate a quick overview every morning and evening. Group chats, DMs, everything. Then it archives that into a local knowledge base.
Why? Because I got tired of the “where did that person send that file?” game. You know the one. You’re writing code, you remember someone sent you a document three days ago, you start scrolling through your IM history, you can’t find it, you ask them, they take an hour to reply. It’s a massive time sink.
With Hermes, I just search the local knowledge base. Done.
Business skills. I tried writing custom skills for different workflows. It’s exhausting. Business processes change constantly. You write a skill, then a week later something shifts and you have to rewrite the whole thing.
Hermes handles that better. It can generate its own skills based on the task. And it self-iterates. I’ve spent way less time maintaining skills since I switched.
Calendar tasks. This is the obvious one. Send messages, schedule meetings, coordinate emails. I’ve turned most of it into cron jobs. But — and this is important — I don’t let it run on YOLO mode. Everything that involves writing gets flagged for approval. It doesn’t take much time and it doubles as a reminder.
Remote tasks. This one’s niche but useful. I have Hermes installed on my girlfriend’s computer too, paired with my IM account. She’s not great with AI tools. Before, she’d send me something, I’d do it, and send it back. Now Hermes just handles it directly on her machine.
One time I was in the bathroom, got pinged about something at work, and used Hermes to fetch the info and reply without leaving the room. I’m not proud of that, but it worked.
What Actually Matters
Here’s the short version:
I tried Cursor — good but expensive and unsustainable.
I tried Trae — didn’t last.
I tried Claude Code — too locked down and didn’t work well with my setup.
I landed on OpenCode — flexible, cheap, and works with whatever model I want.
And Hermes covers everything OpenCode doesn’t.
I could be wrong about Claude Code being overhyped. Maybe the next version changes everything. But for now, this setup works. It’s cheap, it’s fast, and I’m not stuck with one provider.
Also — and this is the part that matters — I don’t think about tools anymore. I just code. That’s the win.