Any LLM. Switch mid-session.

Multi-Model Support

Kognisant works with any LLM, local or cloud. Ollama, llama.cpp, OpenAI, DeepSeek, Groq, NVidia, Kimi, Nebius, or any OpenAI-compatible endpoint. Switch models mid-session with /model. The system tracks per-model reliability, applies token calibration corrections, and auto-switches via circuit breakers when a model fails.

See it in action

How it works

1. 1

   Your model pool is a prioritized list of configured models. The top model is active; others are standby fallbacks.

2. 2

   Circuit breakers monitor each model: 5 failures within 30 seconds opens the breaker (30s cooldown), routing to the next model in the pool.

3. 3

   Token calibration tracks predicted vs actual token counts per model and builds correction factors that improve over time.

4. 4

   Switch mid-session with /model. Context carries over seamlessly between models.

Key capabilities

Use cases

• Use free local Ollama for simple tasks, cloud for complex reasoning
• Auto-recover when a cloud API has downtime
• A/B test different models on the same tasks
• Use reasoning models (deepseek-r1) for architecture decisions

Commands