macOS · Development guide · 5 September 2026

Bring your Spark
to Hermes Agent.

Hermes runs on your Mac. SparkPilot provides its authenticated local model connection and forwards requests to the model running on your DGX Spark.

This guide describes the development integration. We have not published an end-to-end Hermes recording or measured benchmark yet; it is not a verified result for every model or tool version.

Before you start

• A compatible SparkPilot development or beta build on your Mac and a paired Spark with Spark Runtime.
• Hermes Agent installed on the Mac before using the launcher.
• A loaded model that supports the requests your chosen Hermes version makes, including tool calls where needed.

Make the connection

1. Choose your Spark and model. Check that the target is reachable and the model is ready before launching the agent.
2. Open the Mac agent launcher. Choose Hermes Agent in SparkPilot's Launch with Spark screen; SparkPilot starts the installed tool and configures its authenticated local API connection.
3. Keep SparkPilot open. The Mac app serves the local connection while Hermes sends model requests to your Spark.
4. Start with a small, read-only task. Ask Hermes to explain a short function pasted into the conversation, without editing files or running commands.

A repeatable first check

Use a fresh conversation with no private files, web tools or extra integrations enabled and ask the same short question on each run; record the selected model, tool versions and whether the model was already warm.

Explain what this function returns and suggest one test, without editing files or running commands: function double(n) { return n * 2; }

Keep the request and reply visible in the recording, alongside SparkPilot's selected model and the Spark's activity, and distinguish elapsed answer time from any token-per-second figure the runtime actually reports.

Hardware requirement

NVIDIA DGX Spark, specified with 128 GB unified memory

Model and software versions

Record the exact IDs and versions used for your run

Published Hermes measurements

None yet

Your model is local, your agent has its own permissions

Hermes may run commands or contact external services if you enable those capabilities. Review its permissions and security guidance separately; SparkPilot's telemetry promise does not cover third-party tools.

Do not expose the local model endpoint publicly or include API credentials, device addresses or private conversations in a shared recording.

If it does not connect

Check that Hermes is installed, SparkPilot is still open, the paired Spark is reachable and the selected model remains loaded; an OpenAI-compatible connection does not guarantee that every model implements every API or tool-call feature.

NVIDIA also publishes a Hermes Agent playbook with its own vLLM setup and security guidance; this SparkPilot-on-Mac workflow is separate and is not an NVIDIA endorsement.

Follow development updates · Read our privacy commitments