I’ve been using Mistral AI for about a year now. I started using it to have a non-ChatGPT experience for my private needs. On the one hand, I wanted a chat system like ChatGPT; on the other, I cared about privacy. That’s how I dove into the world of AI chat systems beyond the usual suspects.

What have I done with it? I’ve collected ideas for small personal projects, defined my own agents in Le Chat (Mistral AI’s own chat system) —though, to be honest, they’re really just specialized system prompts. Are they truly agents? I have my doubts.
By now, Le Chat has become a constant companion, whether I’m processing private communications, improving and adapting my blog posts (like this one), or just brainstorming. I’ve started optimizing everything around Le Chat: setting up my own projects, organizing libraries. I can safely say that without Le Chat, I wouldn’t be as fast as I am—or want to be.

Then, out of the blue at the start of the year, Mistral Vibe appeared. While Le Chat is Mistral’s chat interface, Vibe is a separate CLI tool designed for coding—similar to Claude Code but with Mistral’s models under the hood. I was hyped—I’ve been using Claude Code constantly and can barely work without it. But my discomfort with depending on an erratic American company pushed me to try Vibe right away.

I’ve tested Mistral Vibe multiple times, tossed it aside, and who knows what else. Last week, there was another update, this time to the model. At first, I felt like an idiot because I somehow missed the devstral-2 model, but I’ve come to appreciate it. In my experience, it’s caught up to Claude—it’s no Opus, but compared to the smaller models, it’s closing the gap.

How does Mistral Vibe work?

Basically, it’s a Python CLI that works similarly to Claude Code. It’s far less feature-rich—no commands, for example. The only thing it supports is skills. How to use them effectively? I’ll get to that later.

Vibe supports the usual modes: plan, work with many follow-up questions (so many), work with few follow-up questions, and a yolo mode (which I now use exclusively—I’m just done with the back-and-forth).

I’ve tried a few things, especially because Claude Code burns through my tokens way too fast. But with Vibe, I’ve never produced anything truly satisfying. Too often, it goes in circles (though the latest model is better at avoiding that), which drove me crazy. What frustrated me most was that I couldn’t use the Get-Shit-Done framework. GSD is a workflow framework that forces structured, step-by-step execution—something I missed in Vibe’s default behavior, which often spiraled into loops. Vibe doesn’t support commands, and GSD relies heavily on them. I wanted to make sure Vibe’s limitations weren’t just due to poor context usage on my part.

So, I waited a few weeks—and now, GSD itself uses skills!

What does GSD bring me in the context of Vibe?

What does GSD bring me in the context of Vibe, besides burning through 27% of my monthly usage in just three days?
I’ve found that it now works much more focused on my instructions. From my experience, the small context window helps, as does the ability to clean the session manually with the clean command.

Why is that important?

I get the impression that Mistral’s models suffer from the same issues as everyone else’s: they remember the beginning and the end of a conversation well, but the middle? That’s where things get fuzzy. So, it’s great to be able to avoid compaction and force a clean instead. With everything properly documented, any model can pick up right where it’s needed.

But I’ve also noticed something I need to verify: Mistral Vibe can be a bit overzealous. Even if the task is just to plan (and you’re not using Vibe’s plan mode), it sometimes jumps straight into solution mode—even when that’s not what you asked for.

What’s my current interim conclusion?

Mistral Vibe is becoming genuinely usable. Even with the minimal GSD skills, it’s significantly more practical than what I’ve seen in the past few months. What wasn’t so great? Well, with Mistral Vibe, compaction can also be your enemy.

I clearly wasn’t paying attention to how full the context was. I let GSD work through a phase, and then—bam—compaction hit. Suddenly, the focus shifted to something completely different than what the phase was supposed to address. The only solution? ABORT! ABORT! ABORT!

pi.dev: The Game Changer

For other reasons, I stumbled upon pi.dev. I was looking for a way to connect to a local Ollama model via MCP—either by building something myself (spoiler: that didn’t work) or by using an existing tool that supports it.
Ollama was my attempt to run models locally for privacy-sensitive tasks, but pi.dev’s GSD integration with Mistral’s remote models turned out to be the better fit for my workflow.
In the end, I landed on pi.dev and was genuinely impressed. My initial test with it was smooth and worked well (maybe another blog post for that). But what’s even more interesting is that pi.dev already has a Get-Shit-Done extension (or package, in pi.dev terms).

So, I installed it, configured Mistral as the remote model, and entered the latest Mistral-Medium-3.5. And just like that, I was sold. It’s just as usable as Claude. I have access to the full GSD feature set with a non-Claude model, and I haven’t noticed any differences. Since I already have to be token-conscious with Claude, Opus is out of the question for me. Mistral Medium 3.5 isn’t any worse, and in my context, it’s still cheaper. Either way, I’m getting much further than I ever did with Claude.

Why Mistral Medium 3.5 + pi.dev + GSD is All I Need

Spoiler: For 90% of development tasks, it is.
Mistral Medium 3.5 + pi.dev + GSD now covers my workflows just as effectively as Claude’s Haiku and Sonnet. The only time I’d reach for Opus is for deep, multi-step reasoning (e.g., designing complex algorithms from scratch). But for refactoring, debugging, or structured coding? Medium 3.5 is more than enough.

The real win? Cost, control, and flexibility.

• Token efficiency: Medium 3.5 lets me iterate 3–4x more than Sonnet for the same budget.
• Workflow flexibility: GSD’s structure + pi.dev’s integration means I’m no longer tied to a single ecosystem. And here’s the kicker: Mistral Medium 3.5 isn’t just great within Vibe—it shines anywhere you can plug it in, making it a true game-changer.
• Future-proofing: Mistral’s rapid improvements (see: devstral-2) suggest the gap with top-tier models will only shrink.

Bottom line: If you’re not building the next LLVM compiler, you probably don’t need Opus. And with this setup, you won’t feel like you’re missing out.

Disclaimer: This blog post was written by me, with the help of Le Chat, an AI assistant by Mistral AI.