nemo-helix

NeMo Helix is a unified platform from NVIDIA that consolidates NeMo agent libraries under a single CLI, Python SDK, and browser-based UI. Its purpose is to make production agents faster, more accurate, and safer. Key capabilities include: - Secure agents: Content safety guardrails, jailbreak detection, PII redaction, anonymization for training data, and automated red-teaming via garak. - Evaluate agents: LLM-as-judge scoring, deterministic and agentic benchmarks, RAG evaluation, and Harbor-backed regression test suites. - Tune agents: Skill optimization, prompt and hyperparameter tuning, and Switchyard model routing for performance improvement. - Build agents: Based on the NVIDIA NeMo Agent Toolkit (NAT) for LangGraph-based agents, with shared infrastructure including an Inference Gateway, Secrets management, File Store, Entity Store, and Jobs service. - Synthetic data: Data Designer for generating training and evaluation data. - NeMo Studio: A browser UI (currently alpha) for chat, monitoring, and reviewing optimization suggestions. The CLI remains the primary interaction surface. Installation is via Python 3.12-3.14 using uv, requiring an API key from a supported inference provider (NVIDIA Build, OpenAI, Anthropic, Google Gemini, or a local Ollama instance). The `nemo setup` command bootstraps local services, registers your LLM provider, discovers models, installs agent skills, and can deploy a sample calculator agent. Coding agents such as Claude Code, Cursor, Codex, and OpenCode are natively supported through installable NeMo skills. Platform services expose a REST API at localhost:8080 for inference and management operations. License: Apache 2.0.