the problem
You’re using Opus
to format a JSON file.
Every developer does it. The frontier model is right there, it’s easy, and it works. But you’re paying Opus rates for work a faster, cheaper model handles just as well. Across a full project — hundreds of calls, dozens of tasks — that habit is expensive.
the proto.type approach
| Architectural reasoning | Hotaru | → | Opus 4.6 |
| Adversarial review | Seori + Kael | → | GPT-5.4 / o3 |
| Code generation | Iseul + Riku | → | GPT-5.4 / Kimi K2 |
| Test coverage | Brynja | → | Gemini 3 Flash |
| Everything | you, probably | → | Opus 4.6 on everything |
Each agent uses the right model for their job. Simple tasks use fast models. Complex reasoning uses powerful ones. The morning report shows what you saved.
model routing
Intelligent model assignment
proto.type automatically assigns each agent to the best model for their role. Any provider, any setup.
providers
Every provider. Intelligently routed.
Works with Anthropic, OpenAI, Google, and dozens more. Automatically assigns each agent to the right model for their task.
monitoring
Stream, thought & MCP views
Watch your team work in real time. See their reasoning. For complex projects, they even build custom tools to understand what they're building.
token intelligence
Context tiering & caching
Each agent receives only the context their role requires. The Mission Object is cached on run start. Prompt caching cuts Hotaru’s cost by up to 80% on long runs.
transparency
Every token accounted for
Every action is logged with tokens used and cost. Full breakdown by agent in every morning report. No surprises.
skill library
Powered by gstack
Agents use gstack — a library of engineering best practices. They cite these practices in debates just like they cite the project plan.
local & open-weight
Air-gapped, local & Hugging Face
Run entirely offline with Ollama, LM Studio, or any local model. Works with the full Hugging Face ecosystem. Available on every tier.
“gstack was instrumental in conceiving this vision. The idea that engineering knowledge could be encoded as a skill library — something agents cite like law — came directly from seeing what Garry built.”