MAQPNA

Telecom operators

Network-operations agents, human-in-the-loop on every change.

An agent that can read alarms can save your NOC hours. One that can push config to the core network needs limits. MAQPNA lets agents diagnose freely, holds changes for an engineer and keeps every action on record.

Why NOC teams keep agents read-only

The blast radius of a wrong change is the whole network.

  • Production changesA misconfigured change can take down service for millions of subscribers.
  • Blast radiusAn agent with broad credentials can reach far more than the task needs.
  • Change controlChanges must follow the change process, with an engineer accountable for each.
  • Subscriber dataSubscriber identifiers and locations must not leak into prompts or logs.

How MAQPNA solves it

Diagnose freely, change carefully

  • Human-in-the-loop change approval

    Rules hold config pushes, restarts and scaling for an engineer, or two for core changes, and record who approved.

  • Per-tool, per-argument policy

    Allow reads everywhere; deny writes to production namespaces; rate-limit everything.

  • Isolation per session

    Sandboxes with default-deny NetworkPolicy contain what a session can reach.

  • Kill switch

    maqpna kill stops an agent across the installation within about a second.

  • DLP

    Redact subscriber identifiers before they reach a model or leave the gateway.

  • Policy testing

    Replay real traffic against a new policy and run policy test suites before rollout.

Architecture

A network-operations agent in the NOC

The agent reads telemetry and tickets freely. Any change to network elements is held for an engineer, and production writes outside the change window are denied.

Operator network · NOC cluster · your keys

Sandboxes

tier-0 (gVisor) or tier-1 (microVM)

  • noc-triage
  • capacity-planner
  • ran-optimiser

MAQPNA gateway

identity · policy · DLP · budgets · approval · audit

  • Allowedoss.get_alarmsread-only-tools
  • Held for approvalnetconf.deploy_configdestructive-needs-human
  • Deniednetconf.patch prod-coreproduction-writes-blocked

Your systems

  • MCPOSS and alarms
  • MCPNETCONF adapter
  • MCPticketing
Hash-chained audit ledger · alerts and SIEM export
Illustrative names. You write your own policies.

Evidence

Evidence for resilience and security reviews

Controls and evidence, mapped. Not a certification.

  • TM Forum autonomous networks

    the 2026 catalyst C26.0.941 argues for "governance by structure, not by prompt". The gateway enforces policy outside the model, so nothing an agent reads or writes in a prompt can loosen it.

  • Closed-loop change approval

    the agent proposes, an engineer approves, and each change records the session, the rule, the approver and the result.

  • NIS2 Art. 21 and 23

    access control, logging and incident-response evidence for agent actions, with a ledger timeline for incident reports.

  • EU AI Act Art. 12 and 14

    record-keeping and human oversight controls. How the EU AI Act applies to agents.

  • Honest scope

    MAQPNA does not judge whether a change is correct; your engineers do.

FAQ

Common questions

What does MAQPNA do for telecom?

An agent that can read alarms can save your NOC hours. One that can push config to the core network needs limits. MAQPNA lets agents diagnose freely, holds changes for an engineer and keeps every action on record.

What stops telecom from putting AI agents into production?

Production changes: A misconfigured change can take down service for millions of subscribers. Blast radius: An agent with broad credentials can reach far more than the task needs. Change control: Changes must follow the change process, with an engineer accountable for each. Subscriber data: Subscriber identifiers and locations must not leak into prompts or logs.

Which controls does MAQPNA enforce?

Human-in-the-loop change approval: Rules hold config pushes, restarts and scaling for an engineer, or two for core changes, and record who approved. Per-tool, per-argument policy: Allow reads everywhere; deny writes to production namespaces; rate-limit everything. Isolation per session: Sandboxes with default-deny NetworkPolicy contain what a session can reach. Kill switch: maqpna kill stops an agent across the installation within about a second. DLP: Redact subscriber identifiers before they reach a model or leave the gateway. Policy testing: Replay real traffic against a new policy and run policy test suites before rollout.

Which regulations does it supply evidence for?

TM Forum autonomous networks: the 2026 catalyst C26.0.941 argues for "governance by structure, not by prompt". The gateway enforces policy outside the model, so nothing an agent reads or writes in a prompt can loosen it. Closed-loop change approval: the agent proposes, an engineer approves, and each change records the session, the rule, the approver and the result. NIS2 Art. 21 and 23: access control, logging and incident-response evidence for agent actions, with a ledger timeline for incident reports. EU AI Act Art. 12 and 14: record-keeping and human oversight controls. How the EU AI Act applies to agents. Honest scope: MAQPNA does not judge whether a change is correct; your engineers do. MAQPNA supplies controls and evidence; it does not certify compliance.

Where does it run?

The agent reads telemetry and tickets freely. Any change to network elements is held for an engineer, and production writes outside the change window are denied. The same runtime runs on a laptop, in your cloud account, on Kubernetes or air-gapped.

How do network-operations agents get human-in-the-loop change approval?

The agent proposes, an engineer approves, and each change records the session, the rule, the approver and the result.

Give your NOC a safe agent.

Policy, approval flow and evidence, walked through with you.

More install options
curl -fsSL https://maqpna.com/install.sh | sh