Nova Transparency Report: How Our AI Research Swarm Burned Through the Monthly API Budget

We owe you an explanation.

Over the past week, some of you may have noticed intermittent slowdowns or fallback responses from the AILinux Nova AI system. Here’s what happened — and what we’ve done about it.

What went wrong

Nova’s MCP (Model Context Protocol) backend runs in autonomous mode — meaning our AI agents continuously scan the codebase, identify improvements, run security audits, and generate research proposals without human intervention. This is by design: it’s how a solo developer operates a system that would normally require an entire team.

The problem: our research swarm got a little too enthusiastic. Over several days, the autonomous agents executed dozens of deep code analysis runs, security audits, and improvement proposals — each consuming significant tokens from our cloud API providers. The result:

• Anthropic Claude API — hit the $50/month spending cap on March 13th, paused until April 1st
• Groq — burned through 97,000 of 100,000 daily tokens
• Gemini — approaching new billing tier caps starting April 2026

In short: our AI agents were so productive that they consumed the entire monthly budget in research swarms alone. The irony is not lost on us.

What we’ve fixed

Rather than just throwing more money at the problem, we rebuilt the infrastructure to be resilient and cost-aware:

• OpenRouter Free-Tier Agents — We deployed 3 new agents running on free 120B+ parameter models (Hermes 405B, Qwen3 Coder, Llama 3.3 70B) via OpenRouter. These handle research, code analysis, and quick tasks at zero cost.
• Budget Guard — A new spending protection system that enforces daily ($2) and monthly ($15) limits on paid API usage. When the budget is exhausted, agents automatically fall back to free models instead of dying.
• Fallback Routing v2.0 — Every agent in the system now has a clear fallback chain that ends at OpenRouter Free. No more cascading failures where one provider going down kills the entire agent network.
• 14 Agents, 5 Running — The system is now fully operational with a mix of paid and free agents, ensuring continuous availability regardless of budget status.

The silver lining

The research swarms that burned through our budget weren’t wasted. They produced dozens of actionable security findings, performance improvements, and code quality proposals — including critical issues like unauthenticated admin endpoints, shell injection vectors, and memory-unsafe upload handlers. These findings are now being systematically addressed by the same agents that discovered them, running on free models.

Lessons learned

When you give autonomous AI agents access to cloud APIs without spending limits, they will use every token available. That’s not a bug — it’s emergent behavior from agents doing exactly what they were designed to do. The fix isn’t to make them less capable, but to give them guardrails.

We’re a one-person operation building what entire teams build. Sometimes that means our AI assistants get ahead of the budget. We’ve fixed the structural problem, and the system is now both more capable and more cost-resilient than before.

Thanks for your patience. Nova is back online and smarter about spending.

— Markus Leitermann, AILinux Project
March 22, 2026