How DeskNexo works, under the hood.
Three components, one chain of trust: a WHMCS plugin that watches tickets, an AI control panel that plans and verifies every action, and your VM platform that executes. Nothing happens that isn't planned, verified twice, and logged forever.
WHMCS Plugin
Installed as a WHMCS addon module. Intercepts tickets via hooks, relays commands to the AI panel, and posts responses back to the customer.
hooks → ticket.created → ingest
api → command relay (HTTPS)
reply → posted back to thread
AI Control Panel
Standalone web interface. Handles LLM processing, RAG lookup, ticket classification, action planning, safety checks, and your dashboard.
llm → planner + verifier
rag → knowledge base retrieval
safety → heuristic + approval gates
VM Platform
Proxmox or VirtFusion API connection. Executes VM actions - reboot, resize, snapshot, firewall, password reset - based on AI directives.
proxmox → api v2 · token-scoped
virtfusion → api · account-scoped
actions → reboot · resize · snapshot
Six steps. Two verifications.
Every ticket follows the same pipeline from arrival to resolution. The flow is deterministic - the AI proposes, the system verifies, and only then does anything execute.
Ticket received
A WHMCS hook ingests the ticket the second it lands. The agent classifies it: info, tech, or escalate.
Read in docs →Context retrieval
RAG pulls the relevant knowledge-base articles, VM history, and previous tickets to ground the answer.
Read in docs →Action planning
The first LLM drafts an action plan: what to do, on which VM, with which parameters.
Read in docs →Dual verification
A second, independent LLM reviews the plan. The heuristic engine scans for dangerous patterns in parallel.
Read in docs →Safety check
Consensus required. If either model disagrees or a heuristic trips, the action is blocked and a human is flagged.
Read in docs →Execute & log
A snapshot is taken first. The action runs, the immutable audit trail is written, and the customer is notified.
Read in docs →Built to be trusted with your infrastructure.
An AI that acts on servers is only useful if it never surprises you. DeskNexo is engineered around one idea: verify twice, snapshot first, and never act alone on something destructive.
Dual LLM verification
Two independent models must agree before any action runs. If they disagree, it is blocked and flagged for review.
Snapshot before every action
A rollback point is taken automatically before touching a VM. If anything goes wrong, it goes back - instantly.
Human approval gates
Destructive operations - deletes, bulk changes, restores - always wait for your explicit approval. No exceptions.
PII masking & GDPR
Customer data is scrubbed before it ever reaches an external LLM. DPAs available, data minimization by design.
Immutable audit trail
Every decision, verification, and action is logged with hash-chain integrity. Trace any change back to its ticket.
Self-hosted LLM option
On Enterprise, run the models on your own infrastructure - data never leaves your network.
The layers that make autonomy safe.
Autonomy without guardrails is just a liability. DeskNexo layers deterministic controls on top of the AI so that trust is engineered, not assumed.
Dual LLM verification
Every action passes through two independent AI systems. The first drafts the action, the second verifies it independently. If they disagree, the action is blocked and flagged for human review.
- First LLM: action planner
- Second LLM: independent verifier
- Consensus required for execution
Heuristic safety layer
A rule-based engine runs parallel to the AI. It validates actions against known dangerous patterns - bulk deletes, cascade operations, unauthorized scope - before any execution.
- Pattern matching on action scope
- Rate-limiting enforcement
- Known attack-vector detection
Human approval gates
Destructive operations never execute automatically. Every delete, bulk operation, or infrastructure-level change requires explicit human approval through a dedicated interface.
- Destructive actions: always manual
- Bulk actions: always manual
- Read operations: automatic with logging
Immutable audit trail
Every decision, action, and verification is logged with hash-chain integrity. Trace any action back to its origin: which ticket triggered it, which models verified it, who approved it.
- Timestamped action logs
- Hash chain for tamper evidence
- Exportable for compliance
Bulk & destructive actions - approval rules
| Action | Approval required | Reason |
|---|---|---|
| Delete VM | Manual (always) | Irreversible, high impact |
| Reboot VM | Auto with logging | Recoverable action |
| Bulk ticket close | Manual (always) | Potential data loss |
| Snapshot restore | Manual (always) | May affect live services |
| Password reset | Auto with logging | Standard operation |
| Resize disk | Manual down · auto up | Downsize is irreversible |
| Firewall rule add | Auto with logging | Easily reversible |
You choose how much autonomy.
DeskNexo scales its own responsibility to match your comfort — from approval-gated actions to fully autonomous resolution.
Copilot
The agent proposes a response and an action plan, then waits for your approval before executing.
Best for: Onboarding, building trust, and sensitive operations.
Autopilot
The agent reads, acts, resolves, and replies autonomously - and notifies you after every action.
Best for: Repetitive L1 issues once trust is established.
Start in Copilot. Scale with Autopilot.
Built for the stack you already run.
No migration, no rewrites. DeskNexo connects to what's in production today - and the plugin architecture is platform-agnostic for what comes next.
WHMCS
Supported nowThe plugin installs as an addon module in under 10 minutes. Hooks intercept tickets, relay commands, and post replies back to the customer thread.
Proxmox VE
Supported nowNative API v2 integration with token-scoped credentials. Full VM lifecycle: status, reboot, resize, snapshots, firewall, password reset.
VirtFusion
Supported nowAccount-scoped API connection with the same action set. Your credentials, your infrastructure - the agent never holds the keys to your host.
Blesta & HostBill
Roadmap · Q4 2026Both billing platforms are in active planning. The plugin architecture is platform-agnostic, so the same agent core maps to either helpdesk.
What the agent can do on your VMs.
Answers grounded in your knowledge.
The agent never guesses. Everything it says and does traces back to your knowledge base, your thresholds, and your rules.
Grounded answers, never invented
Every reply is retrieved from your knowledge base and docs via RAG. If the answer is not in your sources, the agent asks a clarifying question or escalates.
Confidence scoring with thresholds
Each classification and action carries a confidence score. You tune per-category thresholds - below the bar, the ticket goes to a human.
Multi-turn context
The agent keeps thread context across replies, so follow-ups don't restart from zero. Escalation hands over the full conversation, not a summary.
Circuit breaker & locking
Cost anomalies trip a circuit breaker that pauses autonomous mode. A distributed lock prevents two agents from acting on the same ticket or VM.
Questions engineers actually ask.
How does dual LLM verification work exactly?+
Two independent models process every action. The planner drafts the action; the verifier reviews it without seeing the planner's output. Both must agree on the action, the target VM, and the parameters. Disagreement = blocked + flagged for human review.
Can the agent act on anything outside my whitelist?+
No. Actions are restricted to an explicit allow-list you configure per category. Anything outside it is escalated to a human. The heuristic layer also rejects pattern-matched dangerous operations regardless of what the LLM proposes.
Where does customer data go?+
Ticket text is processed by Yeco-PII first so personal data can be detected and redacted. Yeco-PII itself processes that text; after redaction, external LLMs and summary jobs receive redacted text only. Staff do not read conversations. On Enterprise, self-hosted models mean even redacted text never leaves your infrastructure.
Do you train AI on our tickets?+
Yeco-PII is an automated model that must process ticket text to detect and redact personal data (names, emails, IPs, passwords, cards, tokens, and similar). Encrypted storage of original tickets is also processing. After redaction, language models, operational summaries, and (for Beta partners) improvement jobs receive redacted text only. YecoAI staff do not read ticket conversations. Beta opt-in lets us, as controller, use already-redacted DeskNexo ticket text to improve the Service — not unrelated internal chats, email, or infrastructure credentials.
What happens if the agent makes a mistake?+
Every VM action is preceded by an automatic snapshot for one-click rollback. Destructive operations require human approval, and the full decision trail is in the immutable audit log - you can replay exactly what happened and why.
How is the AI panel hosted?+
The control panel runs on our infrastructure (EU) with your VM credentials stored encrypted at rest with rotation policies and scope-limited keys. On Enterprise, the entire panel can be self-hosted alongside the LLM.
How do you prevent cost spikes?+
A circuit breaker monitors token spend and pauses autonomous processing on cost anomalies. You also set hard monthly budgets and receive alerts at 80% and 100% usage.
Does it support Blesta or HostBill?+
WHMCS is supported first. Blesta and HostBill are on the roadmap for Q4 2026 with the same agent core behind a platform-agnostic plugin.
Can I use my own LLM API key?+
Yes, all plans support BYOK. Business and Enterprise include our managed LLM, and Enterprise supports self-hosted models via an OpenAI-compatible API.
Stop answering
the same ticket twice.
Be one of three Beta partners testing DeskNexo in production. Free trial, launch discount, 24/7 support, and certified 24-hour issue resolution - while you shape the product with us.
Limited to three verified hosting providers · WHMCS plugin included · No per-agent seat fees