User-isolated containers
Real process isolation per user — not just tenant separation.
Agentic OS
The OS for AI teams.
Orchestrator, agent roles (Office · Developer · Researcher · SysAdmin · Security Advisor · …), goal engine with dependencies, runners for website and app builds, skill system, multi-LLM router, user-isolated containers — all in one managed runtime.
- 🇪🇺 EU-hosted
- GDPR-native
- EU-AI-Act-ready
- Multi-LLM
- Managed runtime
Multi-LLM Router
ClaudeGPT-4oMistralOpen-Source
How the OS works
From input to execution — one flow, six layers.
Whether a WhatsApp message, voice memo, photo or coding session: everything starts with an input and flows through the same layers of the Agentic OS. Scroll through the stack.
Multi-LLM Router
ClaudeGPT-4oMistralOpen-Source
01 · Input
Every task starts with an input.
WhatsApp message, voice memo, site photo, web form, mobile push, API webhook. Whether app, agentic loop or coding session — the input field is always the beginning. One channel, one event, one task.
- Text · voice · photo · webhook
- WhatsApp · mobile · web · voice
02 · Orchestrator + agent team
One task, multiple roles.
The orchestrator receives the task and spins up the matching agent team: Office, Developer, SysAdmin, Researcher, Security Advisor. Roles hand off to each other, debate interim results, escalate questions. Not a single chatbot doing everything alone.
- Goal engine with dependencies
- Debate & handoff between roles
- + roles grow with new apps
03 · Skills + runners
Capabilities and real infrastructure.
Skills give agents access to email, calendar, WhatsApp, ERP, PDF, custom APIs — tenant-scoped and auditable. Runners are real execution environments: website builds, app deploys, CI/CD pipelines. Not toy tools — productive compute.
- Email · calendar · WhatsApp · Odoo · PDF · ERP · API
- Website runner · app runner · build + CI/CD
04 · Multi-LLM router
Best model per task — no lock-in.
Triage with Haiku, reasoning with Sonnet, vision with GPT-4o, routine with Mistral, code with open source. Each task class gets the right model. If a provider fails or rate-limits, the router fails over automatically.
- Claude · GPT · Mistral · open source · BYOK
- Automatic fallback on outages
05 · Security Gate
Before any action has external impact.
Payments, outgoing mail, ERP writes, price changes: anything MEDIUM or HIGH risk lands at the Security Gate. A human approves, the audit trail is written. LOW risk (show draft, ask for info) passes through.
- Risk score · approval queue · audit trail
- No agent sends blind
06 · User container
One isolated runtime per user.
The whole stack above doesn't run in a shared process — it runs in a container that belongs exclusively to the user. Credentials, interim state, tool state — all physically isolated. Managed runtime: we handle provisioning, scaling, health.
- AES-256-GCM credentials vault · per user
- True process isolation, not just DB filters
8 pillars
The foundation — quotable in eight lines.
The deep-dives above show how it looks. Here's the compact recap, so you have the core at a glance.
Security Gate
Approval workflows for risky actions. Full auditability.
Credentials Vault
AES-256-GCM, per user, per integration. OAuth refresh built in.
Goal engine
Multi-step workflows with dependencies — not single-shot agents.
Multi-LLM router
Best model per task: Claude, GPT, Mistral, open source. No lock-in.
EU-hosted, GDPR-native
Data processing exclusively in Europe. EU-AI-Act-ready.
Cross-system orchestration
One agent, all systems — email, calendar, ERP, industry software.
Auditable
Every AI team action traceable and documented.
Multi-LLM router
Best model per task — no lock-in.
One model per class of task. The AI team picks the right model per task — by quality, cost, latency or hosting requirement. A provider outage automatically reroutes to the next.
| Task | Model family | Why |
|---|---|---|
| Triage & categorisation | Haiku, small models | Cost — 95% of inbox events don't need a large model |
| Decisions & text generation | Sonnet, GPT-4o | Quality on reasoning and customer-facing copy |
| Vision: measurements, damage photos, sketches | Vision-capable LLMs | Multi-modal — text + image in one pass |
| Routine generation, templates | Open-source (Mistral, Llama) | EU hosting, cost, reproducibility for regulated industries |
| Code, tool use, structured outputs | GPT code models | Tool-use performance, JSON-mode stability |
| Long-context recall | Claude 1M context | Large surveyor files, ERP dumps without chunking |
Active providers in the router
- Claude
- GPT-4o / GPT-5
- Mistral
- Llama / DeepSeek
- Haiku
- + BYOK
Every row has a defined fallback. If a provider fails or rate-limits, the task keeps running — just maybe with 0.4s more latency or 20% more token cost. Not a single task fails on a single vendor.
Compliance & trust
European, transparent, verifiable.
Privacy isn't a checkbox. Here's what we commit to and what you can verify.
EU hosting
Compute and storage exclusively in EU data centres. No US cloud in the hot path. Sub-processors documented on the trust page.
GDPR-native
Data processing agreements (DPA) available. Data subject rights (access, deletion, portability) are product features, not compliance theatre.
View DPA →EU-AI-Act-ready
Each app is classified by EU-AI-Act risk class. Transparency logs and human oversight are built into the platform — not retrofitted.
Audit trail
Every AI team action is logged structurally and searchable. Retention configurable per tenant. Export for external audits.
Security policy →Who we are up against
Four camps, one European answer.
The full comparison against legacy enterprise, AI-lab runtimes, consumer EU AI and vertical tools lives on the home page. Here is the short version of why platform buyers come to us.
- Agent-native from day one — not an agent layer on top of a legacy stack (Salesforce, MS, ServiceNow).
- App framework above the agent runtime — not just a raw runtime like Claude Managed Agents or OpenAI Operator.
- Multi-LLM router — no lock-in to a single provider, however large they are.
- Isolated containers per user — real isolation, not just tenant separation.
- EU hosting + GDPR-native — structural, not explained after the fact.
- Open platform — custom apps on equal footing with our example apps.
OpenClaw vs. Linkworld
You know OpenClaw? Linkworld is the business edition.
OpenClaw shows what agentic AI delivers for individuals — 300,000 GitHub stars in months. For companies the governance layer is missing: tenant isolation, approval workflows, GDPR, packaged vertical apps. Linkworld provides exactly that layer.
OpenClaw (open source)
Personal · Self-host- Open-source AI agent framework
- Self-hosted, model-agnostic
- Personal productivity, small teams
- Skill marketplace with supply-chain risks (Cisco, Gartner, Trend Micro)
- Single-tenant per installation
- You carry security, hosting, updates yourself
Linkworld (enterprise layer)
Business · Managed- Hosted SaaS, EU-sovereign — 100+ tools, same idea
- Multi-tenant with isolated user containers
- Security Gate, approval workflows, AES-256 vault
- GDPR-native, EU AI Act-ready
- Packaged vertical apps (field sales, construction, audit) — ready to use
- Audit logs, BYOK, on-prem option
This is what we build on top.
Three example apps show what the platform is built for. Custom apps are next.