Jonathan Solskov

Software Engineer | Rust Enthusiast | Founder of Zirv

[Jonathan Solskov @glubiz]

Building ZirvFitness solo started as a simple fix for my own messy workout tracking. Now every day I’m deep in the code, shipping updates, and leaning hard on AI tools to brainstorm features, clean up logic, and iterate way faster than I could alone.

It’s still early. Just a handful of users so far, and still just me using the app multiple times a day for my lifts and runs, then turning those real sessions into the next improvement.

If you like following indie builders, fitness apps in progress, or seeing how AI actually helps a solo developer ship and refine, stick around. I’ll post the real ups, downs, and process as I keep going.

App’s free to try: https://t.co/KFSQkWNQYH

#IndieHacker #SoloDev #BuildInPublic #AIforDev #FitnessApp

[Jonathan Solskov @glubiz]

The median savings were 16.3% on cost and 11.1% on agent time. One important thing to note about the benchmark is that Zirv spends a little time organising the work before the AI starts. On a small task, that setup time makes up a bigger part of the total. This is medicated by the Jev proxy built into Zirv, but unfortunately it wouldn’t work in the benchmark. Therefore the medians are much lower than real world use.

[Jonathan Solskov @glubiz]

Zirv now has an orchestrator dashboard to follow your agents and how they talk to each other 🤖

Different models and harnesses, shared skills and workflows.

Benchmarked against Claude Code + Superpowers: up to 42% cheaper, 32% less agent time 💸

If you are interested in what I am cooking, or want to contribute please visit https://t.co/Gfb7dGsVnR 🙏

#Zirv #Claude #Codex #Agentic #Programming #Software

[Jonathan Solskov @glubiz]

I did 10 large coding tasks plus 9-step and 22-step coding sessions, 78 runs total, using Sonnet 5 in both setups. The 42% lower cost and 32% less agent time came from the 9-step session with Jev disabled. Those are best-case figures, not averages across the benchmark.

All tasks are committed alongside the benchmarking tool in the GitHub repo 🙌

[Jonathan Solskov @glubiz]

The Zirv meta harness can now create motion graphics using any agent running within Zirv 🎬

Software development: ✅
Animation: ✅

You can check it out here: https://t.co/Gfb7dGsVnR 🔥

I’ll soon be adding instructional GIFs made with this skill to my other project Zirv Fitness: https://t.co/B6yuuLCfDN 💪

[Jonathan Solskov @glubiz]

Just shipped the harness proxy in zirv using TypeSafe's Jev, and wow, this is a game-changer.

Traditional LLMs give you natural language. Jev gives you deterministic decisions with confidence scores in 70-500ms. For an orchestrator that needs to pick models, complexity, and execution topology on the fly, that's exactly what you need.

The kicker? It costs $0.042/MTok and the speed means the session cost is lower due to the fact that the harness fits the complexity and workload and you're not adding latency to every session start. No complex prompt engineering. No hallucinations. Just structured choices grounded in your actual prompts.

For workloads that vary wildly (trivial scripts to multi-agent orchestrations), this solves the real problem: neither orchestrator-everything nor single-agent works. Jev lets you route intelligently without the weight.

Excited to see this pattern spread.

@typesafeai job well done 👏

[Jonathan Solskov @glubiz]

A bit more about how Zirv CLI works. 🛠️

It supports multiple AI harnesses and models. One can coordinate the work while others handle implementation, testing or review.

They share context, memories, skills and workflows, so project knowledge and instructions can carry across agents. Handoffs let another agent pick up ongoing work with the relevant context.

Rollover is another part of this: 🔄

- Usage-based rollover moves work between harnesses as their available usage changes. That includes handing over the orchestrator role when it reaches its limit, so the whole session doesn’t have to wait for a reset.
- Failure-based rollover moves work away from a harness or model when API errors, proxy failures or service degradation prevent it from continuing.

The configured setup is the starting point, but which harness handles the work can change depending on availability and remaining usage.

That’s what I’m working towards with Zirv: less manual coordination when using several AI tools on the same project. 💪

https://t.co/Gfb7dGsVnR

#AI #DevTools #SoftwareDevelopment #BuildInPublic

[Jonathan Solskov @glubiz]

I made a thing…

The Zirv CLI makes Codex and Claude Code work as one team (more harnesses are on the way).

It handles orchestration, shared context, shared skills, shared memories, workflows, handoffs, prevents context rot and a lot more.

https://t.co/AfbxqeZcXs

[Jonathan Solskov @glubiz]

@HungQ95288 @AnthropicAI It is actually possible to get the $ amount spent using OTel and managed settings. Therefore it is not guesswork, it is Anthropic themselfs that report the usage (both token count, and $ equivalent) on the individual on your team plan.

https://t.co/UlNL3zQzAm

[Jonathan Solskov @glubiz]

Did the Claude Team plans 5 hour limit get nerfed?

I used to get ~$200 equivalent in tokens before I hit the cap, now I’m hitting the cap at ~$100. Anyone else experiencing this?

#ClaudeAI @AnthropicAI

[Jonathan Solskov @glubiz]

I got tired of workout trackers that felt bloated or just didn’t match how I actually train. So I built ZirvFitness.

It’s a simple web and iOS app for logging strength and cardio sessions fast, tracking progress over time, and sticking to phased plans without the extra noise.

Beyond the basics it shows you long-term trends from months of data, pings you on new personal records, estimates your 1RM, auto-schedules deloads, and works across iPhone, Apple Watch and the web so everything stays in sync.

I use it every single day (sometimes multiple times) for my own bodybuilding and cardio workouts, and I’m constantly adding features and improving it based on user feedback, and my experiences with the app.

Sign up free, cancel anytime. Pro is just 4.99€ a month or 39.99€ a year if you want the extra stuff.

Check it out: https://t.co/KFSQkWNQYH

#FitnessApp #WorkoutTracker #Bodybuilding #IndieHacker #SoloDev

[Jonathan Solskov @glubiz]

They have created a new incident on the matter.
Follow along here: https://t.co/PBci41yyWD

[Jonathan Solskov @glubiz]

Did @AnthropicAI just remove the free Fable tokens?

[Jonathan Solskov @glubiz]

Completely agree, and well put: architecture is rediscoverable from code, but rejected approaches are gone forever unless you write them down.

What fixed the architecture drift for me was a librarian agent as the final step before every PR. It reconciles the vault against the context of the current session, so documentation updates aren’t a discipline problem anymore.

And of course I had to have a Dennis 🤣

[Jonathan Solskov @glubiz]

I’ve been building ZirvFitness (https://t.co/KFSQkWNQYH), a full-stack fitness tracking app, with Claude as my pair programmer. Some numbers first, because they still surprise me:

• ~260,000 lines of Rust (8 Actix Web microservices + a Dioxus frontend, one Cargo workspace)
• ~103,000 lines of Swift (native iOS app, Watch app, Live Activities, HealthKit)
• Kafka between services, Caddy in front, MySQL underneath, CI/CD that auto-tags and deploys per service
• Shipped by one person

The interesting part isn’t that Claude writes code. It’s what I had to build around Claude to make that sustainable. Three things made the difference:

1. A second brain the agent actually reads.

The repo has an Obsidian vault with 159 pages: architecture maps, a decision log, a work journal, and a “known issues” file of every gotcha that ever cost me a debugging session. Every session starts by reading a system-context file and ends by updating the journal. Claude sessions are stateless; the vault makes the project stateful. When a new session picks up work, it knows what the last one decided and why.

2. Custom subagents as a review board.

I have 14 specialized agents in the repo. Before anything ships, the relevant ones run: one validates Kafka producer/consumer wiring, one audits iOS lifecycle edge cases (foreground/background/lock/kill+resume, the dominant bug class in a fitness tracker), one checks all 7 translation languages for missing keys, one verifies offline-first write paths survive a cold start. My favorite is “dennis”: an adversarial reviewer whose only job is to argue against whatever the other agents approved. Nothing merges until dennis has had a say.

3. Skills that encode process, not just knowledge.

TDD is enforced by a skill: failing test first, no exceptions. There’s a pre-PR checklist skill that catches deployment-blocking issues (version bump cascades, config drift between dev and prod routing). A “verify-before-done” skill that forces actual command output before Claude is allowed to claim something works. The lesson: don’t tell the agent to be careful, give it a procedure.

The mental shift: I stopped treating Claude like autocomplete and started treating it like a team I’m the tech lead of. My job moved up a level: architecture decisions, reviewing diffs, writing down conventions so they’re enforced instead of remembered. The CLAUDE.md files in the repo are basically onboarding docs, except the new hire reads all of them, every time, in seconds.

Is it perfect? No. The guardrails exist because things went wrong: there’s an entry in my docs about a Dockerfile gotcha that broke a deploy, dated and explained, so no future session repeats it. That’s the loop: mistake → documented gotcha → agent that checks for it → never again.

Solo dev in 2026 doesn’t mean writing everything yourself. It means designing the system that designs the system.

[Jonathan Solskov @glubiz]

Started a little side project habit: when something annoys me in daily life, I build an app for it. ZirvFitness (https://t.co/SJ6YqW5uWa) is the first. I got tired of guessing whether my training worked, so now I just track everything and look at the numbers 📈💪