OpenClaw Agent With NeMo Relay

June 4, 2026 · View on GitHub

This experimental NVIDIA NeMo Agent Toolkit example prototypes a primitive agent workflow type for OpenClaw. The workflow invokes openclaw agent, while OpenClaw loads the nemo-relay plugin to observe the agent run and export telemetry.

Integration Flow

flowchart LR
    User["User starts<br/>a run or eval"]

    subgraph Toolkit["NeMo Agent Toolkit"]
        Workflow["Runs the OpenClaw<br/>workflow adapter"]
        Output["Returns results<br/>and eval output"]
    end

    subgraph OpenClaw["OpenClaw"]
        CLI["OpenClaw CLI<br/>runs the agent"]
        Auth["Uses local OpenClaw<br/>configuration"]
        Plugin["Loads the<br/>nemo-relay plugin"]
    end

    subgraph Relay["NeMo Relay"]
        Events["Records agent,<br/>model, and tool events"]
        ATIF["Writes ATIF<br/>trajectory files"]
        OpenInference["Exports OpenInference<br/>spans"]
    end

    subgraph Phoenix["Phoenix"]
        Trace["Displays toolkit<br/>and OpenClaw traces"]
    end

    User --> Workflow --> CLI
    CLI --> Auth
    CLI --> Plugin --> Events
    Events --> ATIF
    Events --> OpenInference --> Trace
    Workflow --> Output
    Workflow -. Phoenix config .-> Trace

NeMo Agent Toolkit owns the workflow and evaluation lifecycle. OpenClaw owns the agent execution. NeMo Relay is loaded through OpenClaw's plugin system, so this workflow does not launch nemo-relay run or import a Relay events.jsonl file. The plugin can write ATIF files and send OpenInference spans directly to Phoenix when those exporters are enabled in OpenClaw config.

Installation And Setup

If you have not already done so, follow the instructions in the Install Guide to create the development environment and install NeMo Agent Toolkit.

Install this workflow package and verify that nat is available:

uv pip install -e examples/experimental/openclaw_agent_adapter
nat --version

Install OpenClaw CLI so the openclaw command is available on PATH, then onboard OpenClaw in the same shell environment that launches nat:

npm install -g openclaw@latest
openclaw --version
openclaw onboard
openclaw doctor

The example configurations use Gateway mode (local: false) because the nemo-relay plugin runs inside OpenClaw Gateway. Configure Gateway for local token-authenticated runs:

openclaw config set gateway.mode local
openclaw config set gateway.bind loopback
openclaw config set gateway.auth.mode token
openclaw config set gateway.auth.token "$(openssl rand -hex 32)"
openclaw config validate

Clone the NeMo Relay source locally, then build and link the NeMo Relay OpenClaw plugin from source:

git clone git@github.com:NVIDIA/NeMo-Relay.git
export NEMO_RELAY_ROOT=/absolute/path/to/NeMo-Relay
(
  cd "$NEMO_RELAY_ROOT"
  npm install --workspace=nemo-relay-node --workspace=nemo-relay-openclaw --ignore-scripts
  npm run build --workspace=nemo-relay-node
  npm run build --workspace=nemo-relay-openclaw
)
openclaw plugins install --link "$NEMO_RELAY_ROOT/integrations/openclaw"

This source install is used because the nemo-relay-openclaw package may not be available from the public npm registry in the test environment.

Enable the nemo-relay plugin and choose the ATIF output directory. Use an absolute path so Gateway writes files to the intended repository checkout even when Gateway starts from a different working directory:

export OPENCLAW_ATIF_DIR="$(pwd)/.tmp/nat-relay-openclaw-atif"
mkdir -p "$OPENCLAW_ATIF_DIR"
openclaw config set 'plugins.entries["nemo-relay"].enabled' true --strict-json
openclaw config set 'plugins.entries["nemo-relay"].hooks.allowConversationAccess' true --strict-json
openclaw config set 'plugins.entries["nemo-relay"].config.plugins.components[0].kind' observability
openclaw config set 'plugins.entries["nemo-relay"].config.plugins.components[0].config.atif.enabled' true --strict-json
openclaw config set 'plugins.entries["nemo-relay"].config.plugins.components[0].config.atif.output_directory' "$OPENCLAW_ATIF_DIR"
openclaw config validate

Start Gateway in a separate terminal and leave it running while you run the workflow:

OPENCLAW_CODEX_APP_SERVER_MODE=guardian \
OPENCLAW_CODEX_APP_SERVER_APPROVAL_POLICY=on-request \
OPENCLAW_CODEX_APP_SERVER_SANDBOX=workspace-write \
openclaw gateway run

In the terminal that runs nat, verify that Gateway loaded the plugin:

openclaw plugins inspect nemo-relay --runtime --json
openclaw gateway call nemoRelay.status --json

Run With NeMo Relay

From the repository root, run the OpenClaw workflow:

nat run \
  --config_file examples/experimental/openclaw_agent_adapter/configs/config-relay.yml \
  --input "Read exactly these files: $(pwd)/examples/experimental/openclaw_agent_adapter/pyproject.toml and $(pwd)/examples/experimental/openclaw_agent_adapter/src/nat_openclaw_agent_adapter/register.py. Summarize how pyproject.toml exposes the nat.components entry point and how register.py registers the _type openclaw_agent workflow with NeMo Agent Toolkit. Do not edit files."

The run should return a normal NeMo Agent Toolkit workflow result summarizing the adapter package and registration code. Because Gateway-side agents normally execute from the OpenClaw workspace, the adapter also passes the configured workflow working_directory as path context and the command above uses absolute file paths.

Phoenix With NeMo Relay

Install the Phoenix integration if it is not already available, then start Phoenix:

uv pip install -e packages/nvidia_nat_phoenix
docker run -it --rm -p 4317:4317 -p 6006:6006 arizephoenix/phoenix:13.22

Enable OpenInference export in the OpenClaw nemo-relay plugin config, then restart Gateway by stopping and rerunning the Gateway command above:

openclaw config set 'plugins.entries["nemo-relay"].config.plugins.components[0].config.openinference.enabled' true --strict-json
openclaw config set 'plugins.entries["nemo-relay"].config.plugins.components[0].config.openinference.endpoint' http://localhost:6006/v1/traces
openclaw config validate

In another terminal, run the Relay/Phoenix config:

nat run \
  --config_file examples/experimental/openclaw_agent_adapter/configs/config-relay-phoenix.yml \
  --input "Read exactly $(pwd)/examples/experimental/openclaw_agent_adapter/README.md and summarize the Run With NeMo Relay section. Do not edit files."

Open http://localhost:6006. The NeMo Agent Toolkit Phoenix config exports the toolkit workflow trace to the nat-relay-openclaw project. The OpenClaw plugin exports OpenClaw agent, model, and tool spans to Phoenix through its OpenInference exporter.

Evaluate With NeMo Relay

The evaluation sample config uses the same OpenClaw plugin setup and measures toolkit workflow runtime. It uses the profiler avg_workflow_runtime evaluator, so install profiler support before running nat eval:

uv pip install -e ".[profiler]"
nat eval \
  --config_file examples/experimental/openclaw_agent_adapter/configs/config-relay-phoenix-eval.yml

Eval outputs are written under ./.tmp/nat/examples/openclaw_agent_adapter/relay_phoenix_eval/. OpenClaw plugin ATIF output is written to the directory configured in OpenClaw.