edgegate-runner

Connect your edge device to EdgeGate — regression gates and a live device registry for on-device AI.

Works on any Linux/macOS box with Python 3.10+: NVIDIA Jetson, Snapdragon-connected hosts, Raspberry Pi, industrial gateways, laptops.

Install

pip install edgegate-runner

Connect a device (60 seconds)

1. In your EdgeGate workspace: Settings → API Keys → Create key (needs Pro tier). Copy the egk_… token.
2. On the device:

export EDGEGATE_TOKEN=egk_...                 # your workspace API key
export EDGEGATE_WORKSPACE_ID=<workspace-uuid> # shown in the dashboard URL / settings

edgegate-runner agent --vendor nvidia --silicon orin-nano-8gb --name my-jetson

The device appears in your workspace dashboard under Device Targets within seconds, with a live/offline badge driven by its 30-second heartbeat. Stop the agent and the badge flips to offline within ~90s.

--vendor is free-form (qualcomm, nvidia, intel, …); --name defaults to the hostname. Add --once for a smoke test that sends a single heartbeat and exits.

Run it as a service (recommended)

# /etc/systemd/system/edgegate-agent.service
[Unit]
Description=EdgeGate device agent (liveness heartbeat)
After=network-online.target
Wants=network-online.target

[Service]
User=<your-user>
EnvironmentFile=/etc/edgegate-agent.env   # EDGEGATE_TOKEN=... EDGEGATE_WORKSPACE_ID=...
ExecStart=/usr/local/bin/edgegate-runner agent --vendor nvidia --name my-jetson
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

sudo systemctl enable --now edgegate-agent

Behavioral Gate on NVIDIA Jetson (llama.cpp)

With pip install llama-cpp-python on the box, the runner executes Behavioral-Gate runs against a GGUF model (vendor nvidia) — same eval sets, same scoring, same signed verdict format as Snapdragon runs. Works online (edgegate-runner run) and air-gapped (edgegate-runner offline).

Behavioral-Gate runs (Snapdragon hosts)

On a host with a Snapdragon device attached via adb, the runner also executes EdgeGate Behavioral-Gate runs — pulling the run config and model bundle, running the gate on-device, and posting back a summary-only signed verdict (raw model output never leaves your box):

edgegate-runner run --run-id <run-id>

Air-gapped? edgegate-runner offline --config config.json runs entirely from local files.

API and workflow endpoints (n8n, Zapier, Make, OpenAI-compatible)

New in 0.2.0. The gate is not limited to on-device models — it can target an HTTP endpoint, so the thing under test can be an n8n workflow or any chat API.

EdgeGate runs these for you by default, with nothing to install. Use the runner here when the endpoint sits on a private network EdgeGate cannot reach, or when raw model output must never leave your infrastructure.

Capture a baseline from your endpoint:

edgegate-runner capture --config capture.json --out reference.json

{
  "eval_set_path": "eval_set.yaml",
  "system_prompt": "You are a support workflow.",
  "decode_config": {},
  "http": {
    "endpoint_url": "https://your-n8n.example.com/webhook/support",
    "transport": "webhook",
    "request_template": {"chatInput": "{{prompt}}", "case_id": "{{case_id}}"},
    "response_text_path": "reply"
  }
}

Upload reference.json as the golden reference to certify the baseline, then gate against it with edgegate-runner run --run-id <run-id>.

• transport is webhook or openai_chat.
• response_text_path is where the reply text lives in the response. Getting it wrong yields empty text, which scores as a refusal and would make every later run pass trivially — so capture refuses outright if every case comes back empty.
• The endpoint credential is read from EDGEGATE_HTTP_API_KEY in the runner's own environment. It is never sent to EdgeGate.
• Create the run with execution: "runner". A run created as hosted is executed by EdgeGate, and the runner is refused it (409).

Configuration

Docs: https://edgegate.frozo.ai/docs · © EdgeGate. Powered by Qualcomm AI Hub.