A virtual office that manages projects, handles support, tracks finances, ensures compliance, and gets smarter every week. You read a monthly report over coffee. That's your job now.
3 agents handle your communication. Messages come in, get sorted, get answered. You see a daily summary of what happened.
7 agents run your operations. Projects tracked. Tickets handled. Quality reviewed. Monthly brief delivered. You manage by exception.
17 agents across 7 departments. Compliance built in. Self-improving. The virtual office knows your business better than a new hire ever could.
Your agents run on NVIDIA's enterprise AI platform. Not cloud AI you rent by the request. A supercomputer under your desk. Your data never leaves the building. 80% of tasks run at zero cost.
80% of tasks run on your own machine at zero cost. Complex decisions get a second opinion from cloud AI on demand. You get the best of both — and the bill proves it.
Processes, pain points, tools you use. We map it. No migration needed — agents connect to your existing systems.
Start with Level 01. Three agents. Running on secure NVIDIA infrastructure. Handling real work within days.
Message them like you'd message a colleague. They classify, execute, report back. Slack and other channels come later.
Built-in learning loop. The system analyzes what worked, what didn't, and improves itself. Week 1: 70% quality. Month 6: 95%. You didn't do anything — it just got better.
The AZ compiles everything: project status, support metrics, financial summary, compliance log, and one-page recommendations. Five minutes of your time.
Built on NVIDIA's enterprise agent platform. Not a science experiment — industrial-grade AI infrastructure.
Your data runs locally. Not in someone's cloud. Not in a training dataset. Under your desk, under your control.
ISO 27001, ISO 42001, ITIL 4 practices built into the agent hierarchy. Not bolted on — baked in.
Every change versioned. Every decision logged. Every action reversible. If something goes wrong, we roll back in seconds.
NVIDIA provides layers 1-2 (hardware + runtime + security). IT.CAFE provides layers 3-5 (platform + agents + intelligence). No custom infrastructure — we build on NVIDIA's enterprise stack.
NVIDIA poskytuje vrstvy 1-2 (hardware + runtime + bezpečnost). IT.CAFE poskytuje vrstvy 3-5 (platforma + agenti + inteligence). Žádná custom infrastruktura — stavíme na enterprise stacku NVIDIA.
Self-improvement engine. Feedback Ledger collects signals from every agent interaction. Weekly analysis cycle proposes skill improvements. Git-versioned, AZ-approved, auto-deployed. The compound intelligence moat.
Motor sebezlepšování. Feedback Ledger sbírá signály z každé interakce agenta. Týdenní analytický cyklus navrhuje vylepšení skills. Verzované v gitu, schválené AZ, automaticky nasazené. Kumulativní příkop inteligence.
Sub-agent orchestration with spawn depth control. Router dispatches to specialist workers. AZ reviews outputs via sampling. Skills defined as versioned markdown files — OpenClaw native format.
Orchestrace sub-agentů s kontrolou hloubky spawnování. Router dispatchuje na specializované workery. AZ reviewuje výstupy vzorkováním. Skills definované jako verzované markdown soubory — nativní formát OpenClaw.
MCP-based tool integration: PostgreSQL + pgvector for persistence, provider-agnostic business tools (Zoho/Odoo/client's own), Git MCP for skill versioning. OpenClaw channels for Telegram/Slack ingress. OpenShell policies for network governance.
MCP integrace nástrojů: PostgreSQL + pgvector pro persistenci, providerově agnostické business nástroje (Zoho/Odoo/vlastní klienta), Git MCP pro verzování skills. OpenClaw kanály pro Telegram/Slack. OpenShell politiky pro síťovou správu.
OpenClaw agent runtime + OpenShell security sandbox. 4-layer kernel-level isolation: network (deny-by-default, hot-reloadable), filesystem (restricted to /sandbox), process (seccomp, no privilege escalation), inference (all model calls routed through gateway). Single command deployment.
OpenClaw agent runtime + OpenShell bezpečnostní sandbox. 4-vrstvá izolace na úrovni kernelu: síť (deny-by-default, hot-reloadable), souborový systém (omezený na /sandbox), procesy (seccomp, bez eskalace práv), inference (všechna volání modelu přes gateway). Nasazení jedním příkazem.
GB10 Grace Blackwell Superchip. Unified memory architecture. Nemotron 3 Super 120B (A12B) via Ollama. Up to 8 concurrent agent sessions per node. Multi-node linking via RoCE 200 GbE for 400B-700B models.
GB10 Grace Blackwell Superchip. Unifikovaná paměťová architektura. Nemotron 3 Super 120B (A12B) přes Ollama. Až 8 současných agentních sessions na uzel. Multi-node propojení přes RoCE 200 GbE pro modely 400B-700B.
Agent behavior defined as markdown files with YAML frontmatter. OpenClaw native format. Hot-reloaded via file watcher. Git-versioned — every change is a commit.
Chování agenta definované jako markdown soubory s YAML frontmatter. Nativní formát OpenClaw. Hot-reload přes file watcher. Verzované v gitu — každá změna je commit.
.agents/skills/*/SKILL.md
Model Context Protocol — standardized tool interface. PostgreSQL, Git, business tools (Zoho/Odoo/Jira). Agents call capabilities, not vendor APIs. Provider-agnostic.
Model Context Protocol — standardizované rozhraní nástrojů. PostgreSQL, Git, business nástroje (Zoho/Odoo/Jira). Agenti volají schopnosti, ne vendor API. Providerově agnostické.
openclaw.json → mcpServers
Hierarchical orchestration via sessions_spawn. Configurable depth (1-5), concurrency limits (1-20 per agent), global cap. Router at depth 0, specialists at depth 1-2.
Hierarchická orchestrace přes sessions_spawn. Konfigurovatelná hloubka (1-5), limity souběžnosti (1-20 na agenta), globální strop. Router na hloubce 0, specialisté na 1-2.
maxSpawnDepth: 2, maxConcurrent: 8
Skill evolution via git worktrees on PGX. Gardener proposes in isolated worktree, AZ reviews diff, merge to main triggers OpenClaw reload. CI/CD for agent intelligence.
Evoluce skills přes git worktrees na PGX. Gardener navrhuje v izolovaném worktree, AZ reviewuje diff, merge do main spustí OpenClaw reload. CI/CD pro agentní inteligenci.
DAI-gardener/ → review → merge → reload
| Layer | Protection | Ochrana | Modifiable | Změnitelné |
|---|---|---|---|---|
| Network | Deny-by-default outbound. Allow specific hosts/ports via policy YAML. | Deny-by-default odchozí. Povolení konkrétních hostů/portů přes YAML politiku. | Hot-reloadable | Hot-reloadable |
| Filesystem | Restricted to /sandbox and /tmp. No host access. | Omezený na /sandbox a /tmp. Žádný přístup k hostu. | Fixed at creation | Fixní při vytvoření |
| Process | Seccomp filters. No privilege escalation. No dangerous syscalls. | Seccomp filtry. Bez eskalace práv. Žádné nebezpečné syscally. | Fixed at creation | Fixní při vytvoření |
| Inference | All LLM calls routed through OpenShell gateway. Agents never contact models directly. | Všechna volání LLM směrována přes OpenShell gateway. Agenti nikdy nekontaktují modely přímo. | Hot-reloadable | Hot-reloadable |
Compliance Agent audits against ISO/IEC 27001 (infosec), ISO/IEC 42001 (AI management), and ITIL 4 (service management) practices. Weekly, monthly, and quarterly audit cycles. Not copyrighted text — operational practices in our own words, mapped to clause references.
Compliance Agent audituje proti postupům ISO/IEC 27001 (infosec), ISO/IEC 42001 (AI management) a ITIL 4 (service management). Týdenní, měsíční a čtvrtletní auditní cykly. Ne kopírovaný text — provozní postupy vlastními slovy, mapované na reference klauzulí.
Each agent has explicit tool allow/deny lists. Router gets sessions_spawn but no file access. Worker gets filesystem but no spawn. Reporter gets database read but no write to business tools. Principle of least privilege, enforced at runtime.
Každý agent má explicitní allow/deny seznamy nástrojů. Router dostane sessions_spawn ale ne přístup k souborům. Worker dostane filesystem ale ne spawn. Reporter dostane čtení databáze ale ne zápis do business nástrojů. Princip nejmenších práv, vynucený za běhu.
| Config | Konfigurace | Model capacity | Kapacita modelu | Concurrent agents | Souběžní agenti | Connectivity | Propojení |
|---|---|---|---|---|---|---|---|
| 1× DGX Spark (GB10) | Up to 120B params | Až 120B parametrů | ~8 | — | |||
| 2× Spark | Up to 400B params | Až 400B parametrů | ~16 | ConnectX / RoCE | |||
| 4× Spark | Up to 700B params | Až 700B parametrů | ~32 | RoCE 200 GbE | |||
| GB300 | Datacenter-class, no ceiling | Datová třída, bez stropu | Unlimited | Bez limitu | NVLink |
Hybrid model: 80% of tasks run locally ($0/task). Remaining 20% route to cloud AI (Claude, GPT-4o) for complex reasoning. Cloud percentage decreases as local hardware scales.
Hybridní model: 80 % úloh běží lokálně ($0/úlohu). Zbylých 20 % směruje do cloud AI (Claude, GPT-4o) pro složitý reasoning. Podíl cloudu klesá s růstem lokálního hardware.
| Component | Komponenta | Technology | Technologie | Status | Stav |
|---|---|---|---|---|---|
| Hardware | NVIDIA DGX Spark (GB10 Grace Blackwell) | RunningBěží | |||
| LLM | Nemotron 3 Super 120B (A12B) via Ollama | RunningBěží | |||
| Agent runtime | Agent runtime | NemoClaw (OpenClaw + OpenShell) | RunningBěží | ||
| Security | Bezpečnost | OpenShell 4-layer sandbox | RunningBěží | ||
| Agents | Agenti | Router, Worker, Reporter (pilot) | DeployingNasazování | ||
| Database | Databáze | PostgreSQL 16 + pgvector (19 tables) | DesignedNavrženo | ||
| Channels | Kanály | Telegram (OpenClaw native) | Config readyKonfigurace připravena | ||
| Business tools | Firemní nástroje | MCP: Zoho / Odoo / client's own | DesignedNavrženo | ||
| Compliance | ISO 27001 + 42001 + ITIL 4 agent | Skill writtenSkill napsán | |||
| PM methodology | PM metodologie | P3.express (OMIMO, CC-BY) | Skill writtenSkill napsán |
We onboard 10 companies this quarter. No fluff call — a live demo of agents running on your actual workflows.
