Skip to main content

Introduction

NVIDIA NemoClaw for LangChain Deep Agents Code runs LangChain Deep Agents Code inside an OpenShell sandbox. The nemo-deepagents command is the Deep Agents-focused alias for NemoClaw. The NemoClaw Deep Agents blueprint brings together LangChain Deep Agents Code, NVIDIA Nemotron 3 Ultra, and the NVIDIA OpenShell runtime so teams can run coding agents with controlled tools, runtime policy, observability, and model routing. Tavily can be enabled for a nemo-deepagents sandbox by applying NemoClaw’s tavily policy preset and registering a tavily-search credential with the OpenShell gateway. The raw Tavily API key stays on the host gateway; it should not be written into the sandbox, project .env files, or Deep Agents config files.

Prerequisites

  • Docker is installed and running.
  • NemoClaw is installed with the LangChain Deep Agents Code agent available.
  • You have a Tavily API key from the Tavily Dashboard.
NemoClaw treats Tavily as an explicit sandbox egress opt-in. Registering the credential alone is not enough; the target sandbox also needs the tavily policy preset.

Setup

If NemoClaw is not installed yet, select the LangChain Deep Agents Code agent before running the installer:
export NEMOCLAW_AGENT=langchain-deepagents-code
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
If NemoClaw is already installed, start Deep Agents onboarding directly:
nemo-deepagents onboard
During onboarding, choose your inference provider, model, sandbox name, and policy tier. The walkthrough uses dcode-tavily as the sandbox name:
nemo-deepagents dcode-tavily status
List the active and available policy presets for the sandbox:
nemo-deepagents dcode-tavily policy-list
Preview the exact egress changes the Tavily preset will apply:
nemo-deepagents dcode-tavily policy-add tavily --dry-run
Apply the maintained tavily preset to the target sandbox:
nemo-deepagents dcode-tavily policy-add tavily --yes
This opens managed Python egress for Tavily search and extract requests to api.tavily.com.
Export your key only in the host shell long enough for NemoClaw to register it:
export TAVILY_API_KEY="tvly-YOUR_API_KEY"
nemo-deepagents credentials add tavily-search --type tavily --credential TAVILY_API_KEY
unset TAVILY_API_KEY
If you need to replace an existing credential, reset it first:
nemo-deepagents credentials reset tavily-search
Rebuild the sandbox so the new policy and gateway provider attach to the runtime:
nemo-deepagents dcode-tavily rebuild
The rebuild should restore the existing workspace state, reapply the policy presets, and restart LangChain Deep Agents Code.
Open the Deep Agents Code terminal:
nemo-deepagents dcode-tavily connect
Ask a current-events question, for example:
What is the latest news on AI?
When Deep Agents requests a web-search action, review the tool call and approve it if the query is appropriate.

How Tavily works in Nemo Deep Agents

LayerWhat it does
tavily policy presetOpens the sandbox Python egress path for Tavily search and extract API calls.
tavily-search gateway credentialStores the Tavily API key on the OpenShell host gateway and injects it at egress time.
Sandbox rebuildAttaches the updated policy and credential provider to the sandbox runtime.
Deep Agents approvalLets you review tool calls before the coding agent performs web search.
Do not place TAVILY_API_KEY in the sandbox, project .env files, /sandbox/.deepagents, or Deep Agents config. NemoClaw’s managed gateway is designed to keep the raw key outside the sandbox.

Good tasks for Nemo Deep Agents + Tavily

  • researching current package, API, and framework behavior before editing code
  • reading and summarizing live documentation during a coding task
  • checking recent release notes, issues, or changelogs
  • grounding implementation choices in current source material
  • extracting clean content from URLs the agent needs to inspect

Troubleshooting

SymptomWhat to check
Web search cannot reach TavilyRun nemo-deepagents dcode-tavily policy-list and confirm tavily is active.
The key is not being usedRe-register tavily-search, then rebuild the sandbox.

Learn more