Set Up the Prerequisite Stack#

The VSG agent needs two services: VSS in search mode, which indexes and searches camera footage, and a Hermes agent server, where the VSG agent runs with a model served by the OpenVINO™ model server. Set up both on one host, with VSS first. If VSS runs on another host, replace <vss-host> with the host address, in the Set Up VSS in Search Mode step.

The VSG agent console calls the Hermes API server on port 8642, which reasons with the Inference Router and OpenVINO model server on the same host and calls tools on the VSS Model Context Protocol (MCP) server

1. Get the Sources#

VSS and the Hermes installer are both in edge-ai-libraries. Step 5 installs the VSG agent skills from the VSG source in metro-ai-suite. Clone both into the same directory:

git clone --filter=blob:none --sparse --branch release-2026.3.0 https://github.com/open-edge-platform/edge-ai-libraries.git
git -C edge-ai-libraries sparse-checkout set sample-applications/video-search-and-summarization frameworks/agent-harness-install
git clone --filter=blob:none --sparse --branch release-2026.3.0 https://github.com/open-edge-platform/metro-ai-suite.git
git -C metro-ai-suite sparse-checkout set virtual-security-guard-agent

The following steps assume you start from this directory.

2. Set Up VSS in Search Mode#

Choose the service credentials and start VSS:

cd edge-ai-libraries/sample-applications/video-search-and-summarization
export REGISTRY_URL=intel TAG=2026.3.0-rc1
export MINIO_ROOT_USER=<user> MINIO_ROOT_PASSWORD=<password>
export POSTGRES_USER=<user> POSTGRES_PASSWORD=<password>
export RABBITMQ_USER=<user> RABBITMQ_PASSWORD=<password>
export MULTIMODAL_EMBEDDING_MODEL="MarketaJu/siglip2-person-description-reid"
export DATAPREP_DETECTION_MODEL="OpenVINO/yolox_s-int8-ov"
source setup.sh --search

Open http://<vss-host>:12345/ to check that VSS is up. VSG and the VSS MCP server both use this address. For GPU embedding, the Milvus vector database backend, proxy settings, and other options, see the VSS Get Started guide.

3. Install the Hermes Agent Server and OpenVINO™ Model Server#

The agent-harness-install installer starts the OpenVINO™ model server and the Inference Router in Docker, installs the Hermes agent server, and connects the agent server to the router. Follow steps 4 through 7 to configure the Hermes agent server for VSG.

Install the Hermes agent server and serve the model:

cd edge-ai-libraries/frameworks/agent-harness-install
export MODEL=OpenVINO/Qwen3.6-35B-A3B-int4-ov
export SERVED_MODEL="${MODEL##*/}"
export HARNESS_OVMS_REST_PORT=8450 && ./install.sh --harness hermes --model "$MODEL"

Note

The VSG agent needs a model that follows a strict reply format and calls tools correctly. Qwen3.6-35B-A3B is a mixture-of-experts model with about 3 billion active parameters, so it is quick but still reliable. Smaller models often break the reply format or misuse tools. Dense models such as OpenVINO/Qwen3.8-27B-int4-ov are more accurate but run slowly.

  • When the installer prompts for “Runtime [host/openshell]”, press Enter to keep host. openshell runs the Hermes agent server in an OpenShell sandbox.

  • When the installer prompts for Host/IP to advertise in printed endpoint URLs, enter the host’s IP address (<HOST_IP>) so the printed URLs work from other machines.

  • The installer installs Docker Engine, Node.js runtime, and the Python interpreter if they are missing, and checks the Intel® GPU. See Get Started for all options.

  • The model is already in OpenVINO™ IR format, so OpenVINO™ model server downloads it without converting it. The GPU, or system memory for an integrated GPU, must hold the model weights, KV cache, and VSS models together.

  • Behind a proxy, export HTTP_PROXY, HTTPS_PROXY and NO_PROXY first. NO_PROXY (and no_proxy) must include localhost, 127.0.0.1, <HOST_IP>, and <vss-host> so that the Hermes agent server reaches the Inference Router and the VSS MCP server directly instead of through the proxy.

Point the Hermes agent server at the router and the model, then check the installation:

hermes config set model.provider custom
hermes config set model.base_url http://<HOST_IP>:8010/v1
hermes config set model.default "$SERVED_MODEL"
./install.sh status

4. Register the VSS MCP Server#

From the installer directory:

./install.sh mcp register vss http://<vss-host>:12345/mcp hermes

This adds mcp_servers.vss to ~/.hermes/config.yaml. The VSS MCP server must be the only entry under mcp_servers. Remove any other entries.

Note

Behind a proxy, the Hermes agent server can reach the VSS MCP server only if the <vss-host> is in NO_PROXY and no_proxy (see step 3). Otherwise, its requests go to the proxy, and the VSG agent cannot search.

5. Install the VSG Agent Skills#

The VSG agent, named Argus, is defined by the files in hermes-skills/:

File

Purpose

SOUL.md

The VSG agent’s identity (the Argus persona) and boundaries. Installed as ~/.hermes/SOUL.md, which applies to every agent on the Hermes host.

virtual-security-guard

The search, validate, trace, and report workflow; the VSS tool playbook, reply format, and guardrails.

vsg-incident-report

Turns confirmed sightings into a report draft.

The Hermes installer also adds bundled skills to ~/.hermes/skills. The Hermes gateway loads every skill in that directory, and the skills compete with the VSG agent for the model’s attention. From the directory you cloned into in step 1, install the VSG agent skills and move all other skills aside:

cd metro-ai-suite/virtual-security-guard-agent
make install-skills INSTALL_ARGS="--exclusive --no-restart"
hermes skills list

The list must show only virtual-security-guard and vsg-incident-report.

  • --exclusive does not delete the other skills. It moves them to ~/.hermes/skills.removed.<timestamp>; move a skill back to restore it.

  • If a different SOUL.md already exists, the installer backs it up as SOUL.md.bak.<timestamp>.

  • --no-restart skips the gateway restart, because the gateway is not running yet. The Hermes gateway loads the skills when you start it in step 7.

Later, run the command again without --no-restart whenever hermes-skills/ changes, and after hermes update if the list shows the bundled skills again. The command restarts the gateway, so start a new Hermes session from the console afterward.

6. Limit the Tools (Optional)#

The console has no login, so anyone who can open it can make the VSG agent use its tools. To give the agent only the tools it needs, run hermes tools, select the API server platform, and keep only skills and the vss MCP tools. Turn off the other toolsets, such as terminal, file, code execution, browser, and web.

7. Enable the API Server and Start the Gateway#

The console connects to the Hermes agent server through the Hermes API server, which runs inside the Hermes gateway. Both are switched off after the installer finishes.

Generate and save the API server key:

openssl rand -hex 32

Important

This key protects the Hermes API server. Use it again as HERMES_API_KEY in VSG’s deploy/.env.

Add these lines to ~/.hermes/.env. Set API_SERVER_KEY to the generated key and API_SERVER_HOST to the Hermes host IP that the VSG host can reach:

API_SERVER_ENABLED=true
API_SERVER_KEY=<the generated key>
API_SERVER_HOST=<hermes-host-ip>

Install the gateway as a user service and start it. It loads the skills from step 5:

hermes gateway install
hermes gateway start
hermes gateway status

Check the API server and the key:

KEY=<the generated key>
curl -s -o /dev/null -w '%{http_code}\n' http://<hermes-host-ip>:8642/health                    # 200
curl -s -o /dev/null -w '%{http_code}\n' http://<hermes-host-ip>:8642/api/sessions?limit=1      # 401
curl -s -o /dev/null -w '%{http_code}\n' -H "Authorization: Bearer $KEY" \
  "http://<hermes-host-ip>:8642/api/sessions?limit=1"                                          # 200

Next Step#

Continue with Get Started. Set HERMES_BASE_URL=http://<hermes-host-ip>:8642 and use the key from step 7 as HERMES_API_KEY.

Remove#

From the VSS directory:

source setup.sh --stop          # stops VSS
source setup.sh --clean-data    # also deletes the VSS data volumes (indexed videos, database); keeps the model caches

From the installer directory:

./uninstall.sh --harness hermes                    # keeps the downloaded model
./uninstall.sh --harness hermes --delete-models    # also deletes the model

This removes the Hermes agent server and its ~/.hermes directory, OpenVINO™ model server, and the router. Add --keep-agent-data to keep ~/.hermes.