Features
CUDA kernel editor with AI autocomplete
Multi-DSL support: CUDA, Triton, CUTE, TileLang, PyTorch, Numba, Mojo
GPU emulator for 86+ GPU architectures
Real-time profiling with NCU integration
Automated kernel benchmarking and comparison
PTX/SASS assembly inspection
Register pressure and race detection
Multi-GPU performance comparison (up to 6 GPUs in Pro)
Custom agents with skills and MCP integrations
Natural language profiling queries
Forge CLI for automatic kernel optimization (CUDA/Triton generation)
Local LLM support (Ollama, vLLM, LM Studio)
Stable SSH remote GPU workflows
Trace agent executions step by step
Monitor real-time dashboards with cost tracking
Online LLM-as-judge evals for quality scoring
Automated insights with unsupervised topic clustering
SmithDB purpose-built for agent traces
Sub-second query performance across millions of traces
Full-text search with inverted index
JSON key-path filtering and trajectory queries
Self-host SmithDB inside your VPC
LangSmith Engine for autonomous issue detection and fixes
Deploy and scale agents with LangSmith Deployment
Sandboxes for safe agent-generated code execution
Fleet agents for no-code agent creation
Supports Python, TypeScript, Go, Java SDKs
