We cover the entire layer: from inference engines (vLLM, TensorRT-LLM) and agent graphs (LangGraph, MCP), through vector databases (Qdrant, BGE-M3) and data streams (ClickHouse, Kafka), to system languages (Rust, Go), orchestration (Kubernetes, Terraform), and GPU (NVIDIA CUDA, H100).
Flagship engines are those we use most frequently in production — the rest are chosen for the specific task, not trends. This breadth allows us to design solutions tailored to the client’s problem, rather than forcing the problem into a single, preferred tool.
We pick models by measurement, not datasheet. The OpenClaw router serves 39 models today — DeepSeek-V4, Mistral Large 3, Qwen3.5/Coder, GLM-5, Gemma 3/4, Devstral-2 and more — each with measured TTFT, throughput and context window. Frontier models (Claude Opus 4, GPT-5, Gemini 3) are integrated when a project requires them. We mask PII before anything reaches the cloud, and compute BGE-M3 embeddings locally — sensitive data never leaves your infrastructure.