NVIDIA-NeMo/nemo-helix

Latest recorded source activity: October 29, 2026 at 22:09

SkillsMP catalog refreshed: November 2, 2026 at 18:02

skills collected: 81

GitHub stars: 132
GitHub forks: 46

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/NVIDIA-NeMo/nemo-helix.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# NeMo Helix ![NEMO Helix](docs/assets/nemo-wordmark.svg) [![CI](https://github.com/NVIDIA-NeMo/nemo-helix/actions/workflows/ci.yaml/badge.svg)](https://github.com/NVIDIA-NeMo/nemo-helix/actions/workflows/ci.yaml) [![License](https://img.shields.io/badge/license-Apache_2.0-D22128?style=flat-square)](LICENSE) [![Python](https://img.shields.io/badge/python-3.12--3.14-3776AB?style=flat-square&logo=python&logoColor=white)](https://www.python.org/) [![Docs](https://img.shields.io/static/v1?label=docs&message=docs.nvidia.com%2Fnemo-helix&color=76B900&style=flat-square&logo=readthedocs&logoColor=white)](https://docs.nvidia.com/nemo-helix) Make the agents you ship faster, more accurate, and safer. NeMo Helix is an open source control plane for improving and hardening production agents. It runs on your infrastructure and plugs into your platform. Evaluate behavior, test changes and measure the results through repeatable workflows. Run Helix locally or on Kubernetes, and access it through REST APIs, a CLI or a Python SDK. Helix also includes NeMo Studio, a lightweight web UI for observability and orchestration. NeMo Helix uses modular plugins for the infrastructure agents need to run at scale. Choose the capabilities you need, from agent execution and sandboxing to file storage, secrets, authentication, model management and inference middleware. Each comes with a default implementation that you can replace with your own infrastructure or tools. ## Get started **Prerequisites:** Python 3.12-3.14, uv, and an API key for an inference provider (OpenRouter, OpenAI, Anthropic, Google Gemini, NVIDIA Build, or a local Ollama instance). Source development needs Git, GNU Make, a C compiler, and either Flox (recommended) or a system toolchain matching `make toolchain-versions`. Docker is required for the control plane's default local operating mode (`nemo setup --start-services` or default `nemo services run`) because deployments default to the Docker executor; remote-deployment, CLI-only, and configs that omit or reconfigure deployments do not require it. See [Local Docker and Executors](docs/get-started/local-docker-and-executors.mdx) for the subprocess vs Docker matrix. Quick install from PyPI: ```bash curl -LsSf https://astral.sh/uv/0.13.4/install.sh | sh export PATH="$HOME/.local/bin:$PATH" uv tool install "nemo-helix[all]" uv tool update-shell # Then restart the shell nemo setup ``` `uv tool install` gives you a global `nemo` command in its own isolated environment, with nothing to activate. The `all` extra adds the Helix services, so `nemo services run` works; without it you get the SDK and CLI only. To import the SDK from your own code, `uv pip install "nemo-helix[all]"` into a virtual environment instead. Source checkout for development: ```bash git clone https://github.com/NVIDIA-NeMo/nemo-helix.git cd nemo-helix # Install Flox first: https://flox.dev/docs/install-flox/install make bootstrap flox -q activate nemo setup ``` `make bootstrap` uses the Flox-pinned uv, Node.js, and pnpm toolchain; it does not require a prior `flox -q activate`. Activate Flox after bootstrap to continue development in the managed environment. Without Flox, install the versions printed by `make toolchain-versions` and a C compiler, then run `make TOOLCHAIN=system bootstrap` followed by `source .venv/bin/activate`. See [SETUP.md](SETUP.md#toolchain-uv-nodejs-pnpm). `nemo setup` starts local services, registers your LLM provider, discovers available models, selects default and fast agent models, installs agent skills, and deploys a sample agent (see more below). Review [Telemetry and Privacy](docs/telemetry-and-privacy.mdx) for the omnibus disclosure covering anonymous telemetry, bundled library telemetry, third-party endpoint notes, and opt-out controls. See **[SETUP.md](SETUP.md)** for the full source setup playbook (local data dir, DB reset, manual service start, troubleshooting). Verify: ```bash nemo services status ``` To permanently reset local state, follow the explicitly confirmed, guarded sequence in [SETUP.md](SETUP.md#question-3--wipe-local-platform-data). It removes the managed ClickHouse container before deleting any bind-mounted data. <details> <summary>Useful CLI commands once setup completes</summary> ```bash nemo --help # All commands nemo models list # Available models nemo chat # Chat with your default model nemo services status # platform health nemo skills list # Skills installed on the platform ``` Every capability is also available via REST API. Model inference uses the model IDs returned from `nemo models list` and is available at: ```text http://localhost:8080/apis/inference-gateway/v2/workspaces/default/openai/-/v1/chat/completions ``` To run Helix services in the foreground in a separate terminal (instead of the background process `nemo setup` starts): ```bash nemo services run ``` </details> <details> <summary>Studio (web UI) bootstrap troubleshooting</summary> If `make bootstrap` reports that Studio asset bootstrap did not complete, the API still runs but the web UI is unavailable until the bundle is built. Ensure Flox is installed, or provide the versions printed by `make toolchain-versions` with `TOOLCHAIN=system`, then run `make bootstrap-studio` from the repository root. </details> <details> <summary>Non-interactive setup (for agents, CI, or scripts)</summary> ```bash export OPENROUTER_API_KEY=sk-or-... export NEMO_DEFAULT_MODEL=default/nvidia-nemotron-3-super-120b-a12b export NEMO_FAST_MODEL="$NEMO_DEFAULT_MODEL" nemo setup --auto --workspace default --start-services --install-skills ``` </details> ## What is in the box ### Capabilities | Domain | What it covers | Components | |---|---|---| | **Connectivity & Observability** | Connect agents to Helix and see what they actually do in production | NeMo Fabric (specialized agents, formerly NAT), NeMo Relay (black-box and general-purpose harnesses), trace ingestion, session and cost inspection | | **Data** | Generate the training and evaluation data you do not have, and keep sensitive data out of it | Data Designer, Safe Synthesizer, Anonymizer| | **Optimization** | Make the agent more accurate, faster or cheaper | Prompt-optimization, Hyper-Parameter Search with Optuna, model-routing with Switchyard| | **Fine-tuning** | Make OSS models more accurate, faster, or cheaper| LoRA, Full SFT, DPO, GRPO for LLMs, Embedding and Re-ranking models with Customizer plugin ( AutoModel, NeMoRL, Unsloth ) | | **Evaluation** | Score candidates against your benchmarks before you ship them | NeMo Gym environments, Harbor eval suites, LLM-as-judge, deterministic, agentic, and RAG metrics, experiments | | **Safety** | Catch unsafe behavior before and during production | Guardrails (content safety, jailbreak detection, PII redaction), garak red-teaming, OpenShell sandboxing | ### Built-in agents Two agents work the loop with you rather than waiting for you to drive it. - **Analyst.** Reads production traces and surfaces where and why the agent is failing. - **Eval Author.** Turns observed behavior into evaluation tasks so the failure does not come back. ### Surfaces Every capability is reachable from every surface. Pick the one that fits the moment. | Surface | Use it for | |---|---| | **NeMo CLI** | The primary interface. Every service under one command, one config, one context. | | **NeMo Skills** | Agent skills installed into Claude Code, Cursor, Codex, or OpenCode, so your coding agent drives Helix in natural language. | | **Python SDK and REST APIs** | Programmatic access for pipelines, services, and products built on Helix. | | **NeMo Studio** | Web UI for chat, job monitoring, optimization & evaluation results, traces, experiments, and governance review. | ### Foundations Shared services every capability builds on: object and data storage, trace ingestion, model registry, experiments, jobs and session management, inference gateway, authentication and authorization, and secrets management. Workspaces and projects scope every artifact, so what your team produces can be shared, referenced, and governed instead of copied between laptops. **These are defaults, not lock-in.** Each foundation is an interface with a working implementation attached, so `nemo setup` gets you running in minutes. When you move to your own environment, swap them: point storage at your object store, the database at your managed Postgres and ClickHouse, identity at your IdP over OIDC, and inference at any OpenAI-compatible gateway. ## Use NeMo Helix from your coding agent After installation, launch your coding agent (Claude Code, Codex, Cursor, OpenCode, etc) from inside the `nemo-helix` directory. This is the primary way of interacting with the NeMo Helix. Things you can ask it to do, once the platform is running: - "Scaffold an agent from this spec and deploy it." - "Fine-tune Nemotron 3.5 Lightining with my dataset, evaluate it and deploy it" - "Run an evaluation against my agent." - "Add content-safety guardrails to my agent." - "Help me optimize my agent." - "Show me what's running on the platform." - "Shut down NeMo cleanly." ## Release notes See the [release notes](https://docs.nvidia.com/nemo-helix/documentation/reference/release-notes) for version-specific features, improvements, and known limitations. ## Skills `nemo setup` installs NeMo skills into your agent of choice, either into the local directory or globally. NeMo Helix-level skills live under `packages/nemo_helix_ext/src/nemo_helix_ext/skills/` and ship with the `nemo-helix` package; plugin-owned skills live under `plugins/<plugin>/src/<plugin>/skills/`. To install or refresh skills for a built-in coding agent, use `--agent`. For another Agent Skills-compatible harness, point `--path` at that harness's skills directory. ```bash nemo skills install --agent claude nemo skills install --agent claude --skill nemo-build-agent --skill nemo-status nemo skills install --path ~/.my-agent/skills nemo setup --install-skills --skills-path ~/.my-agent/skills ``` ## Try the sample agent Interactive `nemo setup` can create a `sample` workspace with a Fabric-based email security agent and evaluation artifacts. Open the Studio link printed at the end of setup to explore them. ## Documentation Full documentation: [NeMo Helix docs](https://docs.nvidia.com/nemo-helix) - [Telemetry and privacy](https://docs.nvidia.com/nemo-helix/documentation/reference/telemetry-and-privacy): anonymous telemetry, data collection, and opt-out controls. - [Setup](https://docs.nvidia.com/nemo-helix/documentation/get-started): installation, providers, SDK. - [CLI reference](https://docs.nvidia.com/nemo-helix/documentation/reference/cli-reference): all commands. - [API reference](https://docs.nvidia.com/nemo-helix/documentation/reference/api-reference): REST endpoints. ## Development See [CONTRIBUTING.md](CONTRIBUTING.md) for development workflow. See [TESTING.md](TESTING.md) for testing strategy. ## License NeMo Helix is licensed under the Apache License 2.0. Third-party open-source dependencies have their own licenses; review them before use.

Loading the repository README…

Skills in this repository

unclassified

Build one pre-cut NeMo Helix release test-scope artifact with feature descriptions, QA validation, stakeholder review fields, and draft release notes from the previous-release-to-release-branch range while verifying release-derived work on main. Use for…

2026-10-21

Showing 40 of 81 collected skills.