How the full Horizon agent works.

Fifteen collectors run on a full-mode agent. The collector set is enforced from the server side as well as in the agent, so a host that should be running a narrower mode cannot accidentally upgrade itself.

• Health. CPU%, memory%, disk%, uptime. The liveness signal underneath every other reading.
• Hardware. CPU model, core count, RAM, motherboard / chassis identifier — read once a day. Lets the portal pick the right TDP profile, recognise hardware refreshes, and pre-populate asset inventories.
• Docker. Container inventory and per-container CPU%. Used by the portal to attribute power draw across containers proportionally and to surface container-level health on the dashboard.
• Network. Interface configuration, listening ports, ARP table, DNS servers, gateway detection, well-known port name mapping. Feeds the CIS topology view and the operational fleet map.
• Filesystem. Mounts, usage percent, inode usage. Metadata only — the agent does not open files or read their contents.
• Software. Installed package inventory with versions (up to 2000 packages). Linux uses dpkg-query or rpm; Windows reads the installed-software registry; macOS reads the equivalent. Feeds the A6.6 unsupported-software check.
• Cybersecurity. Trivy container-image scanning (counts by severity), read-only nmap for network exposure, and a CIS-aligned check suite — firewall state, SSH configuration, update posture, sudo logging, key-based authentication.
• Power. Whole-host electricity draw in watts. Primary counters where available — IPMI on baseboard management controllers, RAPL on x86 Linux, NVIDIA / AMD GPU counters, powermetrics on Apple Silicon, LibreHardwareMonitor on Windows — falling back to a calibrated CPU-time × TDP estimate, labelled accordingly so the method travels with the figure.
• Sensors. Thermal readings (CPU, GPU, NVMe, motherboard) and fan tachometer speeds. Useful for cooling-load calculations and for spotting hosts running hot.
• AI models. Localhost-only HTTP probes of common inference services — Ollama, LM Studio, LocalAI, OpenWebUI — capped at a three-second timeout per probe. Returns model names, parameter counts, and VRAM use per loaded model. The agent does not call any external AI service from your host.
• Proxmox. Read-only API calls to the local Proxmox node (PVEAPIToken auth) to enumerate VMs and LXC containers, node status, per-guest CPU / RAM / disk. Excluded from Cyber and ESG agents.
• Proxmox Backup Server. Datastore inventory, recent backup task history, version. Separate API token format (PBSAPIToken). Lets the portal show backup success rates next to the workloads they protect.
• Logs. System log headlines only — ERROR / WARNING / CRITICAL severity, capped at 200 entries and 200 KB per push. Sources are journald on Linux, the Windows Event Log, and syslog on macOS / BSD. The agent does not stream verbose application logs and does not send body text beyond the severity-filtered headline.
• Cyber compliance. CE+ A4–A8 self-assessment evidence: A4 (firewalls), A5 (secure configuration), A6 (security updates), A7 (user access), A8 (malware protection). The same collector the Cyber-mode agent uses.
• Hermes. Detects the presence of the Hermes AI framework on the host — CLI binary, data directory, running process — and reads its public configuration: version, inference provider and model, workspace path, skills inventory, cron jobs, memory metadata. It is a presence and configuration read, not a content read; Hermes project files and conversations are not transmitted.

The full agent is broad — but breadth is in the metadata layer, not in the content layer. The agent does not read documents, conversations, or user secrets; it reads the shape of the system the host is running.

• No file contents. The filesystem collector reports mount points and capacity, never the bytes inside them.
• No full log content. Logs are severity-filtered headlines, capped at 200 entries and 200 KB per push. Application logs, request bodies, and verbose journald output are not transmitted.
• No process command lines or environment variables. The agent knows a process exists and how much CPU and memory it is using; it does not read the arguments it was invoked with or the environment it inherited.
• No browser history, cached credentials, or stored secrets.
• No user account contents. Account existence and permission groups are visible at the OS level; passwords, tokens, and session data are not.
• No Hermes project content. The Hermes collector reads framework presence and configuration. Conversation history, project files, and skill outputs are not transmitted.
• No AI inference traffic. The AI-models collector is a localhost HTTP probe — it asks the inference server what it has loaded; it does not read prompts, completions, or model weights.
• No silent remote control. Remote-support sessions are initiated from the portal, audit-logged, and visible to the operator on the host while in progress. There is no covert command channel.

Deployment is the same three-step shape as the narrower agents. The mode is selected when the install token is generated, and the collector set is enforced server-side so a host cannot accidentally drift into a wider mode than it was provisioned for.

1. Generate a token in the portal. Pick the company, choose "Horizon mode", copy the install command.
2. Run the install command on each host. A single line on Windows (PowerShell as admin), macOS, or Linux. The installer drops the agent into a known path, registers it as a service or LaunchDaemon, and exits.
3. Wait a minute. The agent registers itself with the portal on first run; the first telemetry batch lands within sixty seconds. Tabs (Threats, ESG Reporting, CE+ Controls, AI Models, PBS) populate as their underlying collectors return data.

Hosts can be moved between modes by issuing a new install token — the change takes effect on the next check-in, no re-install required.