IOTA Qantara · قنطرة · the arch bridge

Use AI everywhere. Keep control of it.

Qantara is the on-premises AI gateway for banks, insurers, authorities and enterprises. Every model call and every tool call crosses one governed bridge, inside your perimeter.

One endpoint

Compatible with the OpenAI and Anthropic APIs. Adopting Qantara is a base URL and a key.

Inside your perimeter

Your Kubernetes, air-gapped if you need it. Nothing calls home.

Evidence by default

Spend in a ledger, decisions in your SIEM, governance in a tamper-evident log.

Works with the models and platforms you already run

OpenAI-compatible APIAnthropic-compatible APIOn-premises vLLMAzureBedrockVertexOCIIBM watsonxKubernetesOpenShiftAKSPostgreSQLOIDCSCIMSIEM over OTLPPrometheus

What boards and regulators now ask

Three questions most organisations cannot answer today.

  1. Who is using which AI?

    Which teams, applications and agents call which models and tools, and with whose approval.

  2. What is it costing?

    Spend by team, by model and by use case, against a budget someone owns.

  3. Where does our data go?

    Which personal and classified data reached which model, in which country.

Each answer should take minutes and come with evidence. Qantara makes that the default.

Why it is hard today

The usual answers leave a gap a regulator will find.

AI governance is an infrastructure problem before it is a model problem.

The usual answerWhat it leaves open
Block AIStaff use it anyway, outside any control. Shadow AI grows.
One approved chatbotAgents, coding tools and applications still call models directly.
A cloud AI gatewayPrompts and logs leave the perimeter. Hard to defend for classified data.
Build it in-houseMonths of work before the first control, then years of upkeep.

What Qantara is

One on-premises control plane that every AI interaction crosses.

Qantara is the Arabic word for an arch bridge. It is the single point every AI interaction in the organisation crosses: model traffic, tool traffic, and the governance around both.

Illustrative console view with sample data.

Model traffic

AI Gateway

One endpoint for every model. Routing, keys, budgets, guardrails and audit.

Tool traffic

MCP Gateway

One endpoint for every tool an agent uses. Identity and policy on each call.

Catalogue

MCP Registry

The approved list of tools and agents, synced offline for air-gapped sites.

Execution

MCP Runtime

Each tool server in its own hardened container, with no ambient credentials.

One installation per customer, on the customer’s Kubernetes. No shared SaaS plane. No runtime internet access needed.

Two gateways, one identity, one ledger.

  • Applications keep their SDKs. They change a base URL and a key, and nothing else.
  • One sign-in governs both model traffic and tool traffic.
  • Every billed request lands in the ledger, and every decision goes to your SIEM.

The full architecture is in the product deck, available on request.

Capabilities

What crosses the bridge, and what happens on the way.

Models

One endpoint for every model, on-premises or hosted.

Every provider you use

On-premises vLLM, Azure, Bedrock, Vertex, OCI, IBM watsonx and regional hosted APIs.

No client changes

OpenAI- and Anthropic-compatible APIs, streaming included.

Routing that thinks

Semantic routing by request, load-aware routing across self-hosted models, and fallback when a backend fails.

Keys per team and app

Virtual API keys, hashed and expiring, bound to teams, models and source networks.

Residency enforced

Cloud backends admitted only in allowlisted regions. In-Kingdom by configuration, not by promise.

Response cache

Exact and semantic caching, with prompt-cache pricing counted correctly in spend.

Spend

Every token priced, every request billed once.

A ledger records every billed request with tokens, cost and timing. Budgets are read from it in memory, so a hard stop costs no latency.

  • Budgets per organisation, team, key and model, with hard stops
  • Per-minute request and token limits
  • Alerts, forecasts and time-bound budget boosts
  • Chargeback reports and statements per team
  • Usage drill-down by team, user, model and day

429

Budget exhausted: the call stops, with a clear reason.

403

Model not allowed for this team or key.

1×

Each request carries an id the ledger writes once. No double billing on retry.

0

Rows missing or extra under sustained load in IOTA’s billing tests.

Data protection

Sensitive data is caught before a model sees it.

Guardrails both ways

Secrets, personal data, national IDs and card numbers masked or blocked in prompts. Prompt injection blocked.

Saudi and GCC aware

Personal-data detection for Saudi and GCC identifiers, with Arabic entity recognition.

Presets per regime

Ready presets for Saudi PDPL, UAE PDPL, GDPR, PCI DSS and agent safety.

Classified data stays put

Each application carries an NDMO data class. Its keys reach only backends cleared for that class.

Audit without content

Guardrail events record types and counts, never the content itself.

Retention you set

Verbose capture only for the callers you choose, kept 24 hours by default.

The honest limit: pattern detectors find data with a recognisable shape. Guardrails reduce exposure. They do not promise it is zero.

Agents and tools

Agents get the tools they need, and nothing more.

MCP governance

Every tool call is authenticated, limited to the team’s approved tools, budgeted, inspected by guardrails in both directions and metered.

Agent-to-agent governance

Messages between agents follow the same rules: identity, allowed agents per team, budgets and guardrails.

Your REST APIs as tools

Qantara turns an OpenAPI description into an MCP server it hosts itself. No new code in the source system.

Agent personas and runs

Agents act under their own identity and token, and each run is recorded end to end, with its cost.

Governance

The controls an auditor asks about, built in.

AI inventory

Every AI application registered with owner, purpose, risk tier and data class. Its keys work only while approved.

Maker-checker

Changes to policies, guardrails, budgets, grants and approvals need a second administrator.

Model lifecycle

Models approved, deprecated with a sunset date, or blocked.

Emergency stop

Any model, tool, agent or all traffic stopped for a team or for everyone within seconds.

Access reviews

Unused, non-expiring or overdue keys flagged, then certified or revoked.

Governance log

Admin changes and approvals hash-chained, verifiable on demand.

Policy as code

Policies kept in version control, tried in a simulator before they apply.

Visibility

See who uses which AI, for what, at what cost.

Usage and insights

Drill down by organisation, team, user, model and day. Top lists and activity at a glance.

Live traffic

Each request as it happens, with decision, latency and cost.

Shadow AI discovery

Finds applications that call AI providers directly, around the gateway.

Audit to your SIEM

Every decision and admin change, with actor and scope, delivered durably over OTLP.

Health and metrics

Latency, errors, denials and audit backlog in the console, Prometheus and alerts.

Arabic console

The whole console in Arabic and English, right to left where it should be.

What each leader gets

One platform, a different answer for each seat at the table.

CIO and CTO

Choice without sprawl

Any model, on-premises or hosted, behind one endpoint. Switch models without touching applications.

CISO

Data stays where it should

Personal and classified data masked or blocked before it reaches a model. Every decision in the SIEM.

CFO

Spend you can explain

Budgets with hard stops, forecasts and chargeback per team. Every request billed once.

Risk and compliance

Evidence on demand

AI inventory, maker-checker approvals, access reviews and an emergency stop, mapped to SAMA, NCA, NDMO and PDPL.

Sovereign by default

Nothing leaves. Not the prompts, not the logs, not the licence check.

Built to run where the internet does not reach.

  • Offline bundle

    Every image and chart in one signed bundle. One script verifies it, mirrors it into your registry and installs.

  • Signed and listed

    Images and charts signed by IOTA, with SBOMs, build provenance and every image pinned by digest.

  • Offline licence

    The licence is requested, issued and installed as files. It is signed and bound to the installation.

  • Your Kubernetes

    OpenShift, AKS or upstream Kubernetes. One Helm chart. Your PostgreSQL, your identity provider, your SIEM.

  • Safe upgrades

    A maintenance window, an atomic upgrade, post-deployment checks and automatic rollback in one command.

  • No egress

    Models served in-cluster, catalogues synced offline. Nothing calls home.

Cloud models stay available where policy allows, limited to the regions you approve.

Measured, not promised

Governance that does not slow the model down.

1.5 ms

added at p50 by the full governed path

10,938

requests a second on one replica, none failed

5,000

concurrent streaming sessions, none failed

0

failed requests in a rolling restart, a replica loss or a 20-second database outage

Exact billing

After every failure test, no request was lost and none was billed twice.

IOTA tests, November 2026: a local Kubernetes cluster on a single 12-core machine, with a mock model so the gateway itself is measured. Every request passed through identity, policy, limits, budgets, guardrails and the ledger. A floor, not a ceiling. A pilot on your hardware confirms the numbers for you.

Regulatory alignment

Mapped to the regimes you answer to.

Supports a customer’s own assessment. Not a certification and not legal or compliance advice.

RegimeWhere Qantara supports it
Saudi PDPLPersonal data masked, in-Kingdom regions only, access control, records
NDMO standardsData classification enforced per backend, audit, residency
NCA ECCIdentity and access, cryptography, event logs, third-party and cloud
SAMA CSFAccess control, segregation of duties, access review, incident response
SDAIA AI EthicsPrivacy, accountability, transparency, agent tool use
Also mappedUAE PDPL, GDPR, EU AI Act, PCI DSS, ISO/IEC 42001 and 27001, OWASP LLM Top 10

Editions and pricing

Start with control. Add governance as you grow.

Core

Every installation

Route and control

All providers, virtual keys, budgets with hard stops, model allowlists, rate limits, audit trail, console and metrics.

Enterprise

Core plus

Operate at scale

Guardrails, team administrators, SCIM, self-service, alerts, chargeback, response cache, policy as code, agent runs and live traffic.

Sovereign

Enterprise plus

Govern for regulators

Saudi and GCC personal data, MCP and agent governance, Shadow AI discovery, semantic routing and cache, forecasts, maker-checker, AI inventory, lifecycle, access reviews and the governance log.

Licensed per installation with named-user seats. Add-ons available. Usage limits warn and never refuse traffic.

Pricing

Enterprise grade only, for now.

Qantara is sold today as an enterprise-grade, licensed product: Core, Enterprise and Sovereign, priced per installation and seat band. Tell us your scope and we will send pricing.

Ask about pricing

Q4 2027

Free and open source

A free, open-source edition of Qantara will be made available in the fourth quarter of 2027.

How we deliver

From first call to production, in four steps.

  1. Assess and architect

    Map AI use, data classes and models. Agree success measures and the pilot scope.

  2. Design for compliance

    Policies, guardrail presets, budgets and approvals set against your regulators’ controls.

  3. Build and integrate

    Install on your Kubernetes, connect identity, SIEM and models. Pilot with real teams.

  4. Run and improve

    Upgrades with checks and rollback, support, and skills handover to your team.

The team that builds the installation also runs it.

Why IOTA

Built in Riyadh. Engineered for regulated business.

2012

established, headquartered in Riyadh

95%

of revenue from regulated industries

Security by default

Hardening and compliance built in from day one.

Sovereign by default

On-premises and air-gapped, so data, models and IP stay inside your perimeter.

Built to be operated

The teams that build systems also run and improve them.

Riyadh · Dubai · London · engineering centres in Chennai and Hyderabad
IOTA Labs is the product arm of IOTA Technologies.

Next steps

See it on your own models.

A pilot runs on your Kubernetes, air-gapped if you need it, with your identity provider and your models. We agree the success measures first.

  1. A discovery call: your AI estate, data classes and regulators
  2. A live demo built around the three questions
  3. A scoped pilot on your Kubernetes, with agreed success measures

Qantara is open to design-partner conversations and early access.

Ask for a demo

Organisations only. Personal addresses such as Gmail or Yahoo are not accepted.
Privacy Policy