I have been building multi model tools for a while. Then cursor released thier multi model feature set last week. Im not paying 20$ a month on top of what i already pay. This is api based but then z.ai and deepseek are cheaper than subscriptions for a good coding assistant and frontier models can do the planning.

Kondi-chat routes between Claude, GPT, DeepSeek, Gemini, Grok, and local Ollama models automatically. It classifies your intent (coding task vs discussion), checks your budget profile, and picks the best available model. Three built-in profiles: `quality` uses frontier models with a $10 cap, `balanced` is the default at $3, and `cheap` runs everything on the fastest/cheapest models for $0.75.

The part I'm most proud of is council deliberation — for hard architectural decisions, you can spawn a multi-model debate where 3-5 models argue across several rounds and a manager synthesizes the final recommendation. It's expensive but it catches things no single model would.

Other stuff that might interest HN: Rust TUI (Ratatui, inline viewport so terminal scroll/select/copy work natively), agent loop with file tools, MCP support, git integration with undo/checkpoints, session resume, per-model cost analytics. MIT licensed.

Tech stack: Node.js backend (TypeScript) + Rust TUI, communicating over JSON-RPC on stdio. The router has three tiers: a small NN trained on your usage data, an LLM-based intent classifier for cold start, and a deterministic rule fallback.

Would love feedback on the routing approach — is automatic model selection actually useful or do people prefer explicit control? Also, Would love some pointers on getting the subscription based models in here.