In Short

Infinito.Nexus 14.0 enables companies to operate AI on their own infrastructure, connect it securely to existing business applications and deploy AI agents such as Hermes and OpenClaw as virtual employees. Sensitive information can be processed by local models, while optional connections to OpenAI, Anthropic and OpenRouter provide access to external frontier models when greater capabilities are needed. Companies decide which data may leave their infrastructure, which tools an agent can use and which actions remain restricted.

Infinito.Nexus 14.0 brings models, enterprise applications and autonomous agents together in one controlled infrastructure. Organizations can now combine local AI for sensitive data with external frontier models and deploy agents such as Hermes and OpenClaw as isolated virtual employees.

Generative AI has already changed how employees search, write, analyze and develop. But most organizations still face an uncomfortable choice: either send business data to an external AI provider or accept the operational complexity of building a local AI platform from scratch.

Infinito.Nexus 14.0 introduces a third option: a hybrid AI architecture in which organizations decide where every workload runs.

Sensitive prompts can be processed by local models through Ollama or LM Studio. Tasks that require the capabilities of frontier models can be routed to OpenAI, Anthropic or OpenRouter. Applications and agents use one centrally managed gateway instead of implementing separate provider integrations.

The result is not merely another enterprise chatbot. Version 14.0 provides the foundation for a secure digital workforce.

One Gateway for Local and Frontier Models

At the center of the new architecture is a shared LiteLLM gateway. It exposes one OpenAI compatible interface to applications and agents while routing requests to the model selected by the organization.

The gateway can publish models from:

• Ollama running entirely on the organization’s own infrastructure
• LM Studio operating as a headless local model server
• OpenAI’s API and the frontier model family familiar from ChatGPT
• Anthropic models
• OpenRouter, providing access to models from multiple external providers

External routes appear only when the corresponding API key has been configured. Each consuming application receives its own virtual gateway key instead of sharing one global credential. This makes individual consumers identifiable and independently revocable.

The optional LiteLLM administration interface is integrated with Keycloak. Administrators can manage models, virtual keys, budgets, rate limits and usage from a central location. A separate emergency login remains available.

For companies, this creates a practical model portfolio instead of another provider dependency.

Route	Best suited for	Data consideration
Local Ollama or LM Studio	Confidential documents, internal knowledge and regulated workflows	Prompts and model processing remain inside the deployment
Direct OpenAI or Anthropic	Complex reasoning and demanding frontier model workloads	Selected information is processed by the external provider
OpenRouter	Flexible model choice, provider comparison and routing	Processing occurs through selected external providers

This distinction matters. Hybrid AI should not mean that sensitive data silently leaves the company whenever a local model is insufficient. It should mean that routing is an explicit architectural decision based on data classification, cost, latency and task complexity.

AI Inside the Applications Employees Already Use

Infinito.Nexus 14.0 integrates the gateway into the existing application landscape. Sixteen applications can use centrally routed AI capabilities instead of communicating with an AI vendor independently.

These include:

• Nextcloud
• Mattermost Agents
• n8n
• Flowise
• Open WebUI
• Moodle
• WordPress AI Engine
• MediaWiki AIEditingAssistant
• the XWiki LLM extension
• Discourse AI
• Zammad Smart Assist
• the Matrix ChatGPT bridge

Each application receives a dedicated gateway credential. The deployment verifies that its configured endpoint remains inside the infrastructure and that the selected model can answer.

This gives organizations a consistent AI layer across collaboration, knowledge management, customer service, automation and content production. Provider selection and credential rotation happen centrally instead of inside every application.

An organization could, for example, process confidential Nextcloud documents with a local model while routing a public market research task to an external frontier model. The employee continues working in the same application while the platform controls the route.

From Chatbots to Virtual Employees

A chatbot waits for a question. An agent can receive a goal, maintain context, consult tools and perform a sequence of actions.

Version 14.0 introduces dedicated roles for two open source agent platforms:

• Hermes Agent from Nous Research
• OpenClaw, an agent with persistent memory, a workspace, browser capabilities and a web based Control UI

Both are connected to the central model gateway and automatically discover the business tools made available by the deployment. Their interfaces are placed behind the Infinito.Nexus reverse proxy and protected through Keycloak Single Sign On.

This allows companies to treat agents as virtual employees with a controlled technical identity:

• They use their own platform account instead of borrowing an employee’s identity.
• They receive their own model gateway credential.
• Their memory and workspace remain in managed persistent storage.
• Their accessible tools are declared by the platform.
• Their connectivity is tested during deployment.
• Their state can participate in the existing backup architecture.

The following report from inside Bracket22, a trading firm built around AI agents, offers a useful illustration of this organizational model. It is not an Infinito.Nexus product demonstration. It shows what becomes possible when agents are treated as operational team members instead of isolated chat windows.

In a conventional business, the same principle could be applied to specialized virtual roles:

• A research agent reads approved knowledge sources and prepares briefings.
• A service agent analyzes support requests and consults Zammad.
• An engineering agent examines repositories and issues in Gitea or GitLab.
• An operations agent reads monitoring information from Checkmk and Prometheus.
• A content agent researches material and prepares WordPress drafts.
• An internal assistant retrieves documents from Nextcloud and coordinates workflows through n8n.
• A facilities agent reads approved Home Assistant entities.

The objective is not to give an autonomous system unrestricted access. The objective is to give it a defined role, a dedicated identity and the minimum tools required for that role.

MCP: The Tool Layer for the Digital Workforce

Models provide reasoning. Tools provide access to the business.

Infinito.Nexus 14.0 uses the Model Context Protocol, or MCP, to connect agents and AI interfaces with enterprise applications. Around twenty applications can now participate in this tool layer.

Native or adapted MCP capabilities are available for systems including:

• Gitea and GitLab
• Nextcloud
• Mattermost
• n8n
• Baserow
• Jenkins
• Home Assistant
• Moodle
• WordPress
• Zammad
• Checkmk
• Prometheus
• Snipe IT
• Jellyfin
• Listmonk
• Fider
• LibreTranslate
• Qdrant

Hermes, OpenClaw and Open WebUI automatically discover compatible MCP services in their deployment. Flowise can also participate as an MCP client.

The platform’s MCP adapter places a controlled boundary in front of applications that expose a useful API but do not provide a sufficiently protected MCP server themselves. Each adapter instance receives only one provider credential and one checked contract.

The release deliberately begins with read only access. Mutating operations are withheld, and reader and writer permissions remain separate. Contracts restrict callable operations, request and response sizes, concurrency, timeouts, pagination and result counts. Unexpected schema changes cause an adapter to fail closed instead of exposing an unreviewed interface.

For a deeper introduction to the protocol itself, this presentation from Anthropic explains how MCP connects AI applications to external data and tools.

Isolation for Agents That Execute Work

An agent with tools is more powerful than a chatbot and therefore requires a stronger runtime boundary.

Hermes and OpenClaw run through the new svc-virt-kata isolation layer. Where supported, the platform can use Kata Containers. Otherwise it installs and selects gVisor, which intercepts container system calls through a userspace kernel.

In Compose deployments, the isolating runtime is assigned directly to the agent service. In Docker Swarm, sandboxed agents are scheduled onto specially labelled nodes whose default container runtime provides the isolation boundary.

The deployment refuses to run these agents as ordinary runc containers. It also prevents configurations that would undermine isolation, including privileged mode, host networking, shared host namespaces and access to the host Docker socket.

This does not make autonomous agents risk free. It does, however, turn isolation from a recommendation into a deployment property that can be tested.

What Local AI Changes for Companies

Local inference is often discussed only in terms of privacy. Its strategic value is broader.

First, it creates a controllable data boundary. Internal documents, customer records, source code and operational telemetry can be processed without sending the prompt to an external model provider.

Second, it gives organizations continuity. A workflow built against the shared gateway is not coupled to one vendor’s endpoint. Models can be replaced or rerouted centrally without rewriting every connected application.

Third, it creates a path for controlled autonomy. Agents can work with real systems through narrowly defined tool contracts instead of receiving broad network access or human administrator credentials.

Fourth, it makes costs more governable. Routine classification, summarization and retrieval can run locally. Expensive frontier models can be reserved for tasks where their additional capabilities create measurable value.

Finally, it brings AI under the same operational discipline as the rest of the infrastructure: identity management, deployment automation, health checks, backup, diagnostics and access control.

Security Is a Routing Decision

No platform can make every AI workload private simply by offering a local model.

When an organization selects an external OpenAI, Anthropic or OpenRouter route, the information included in that request leaves the local deployment. When a local Ollama or LM Studio route is selected, model processing remains inside the infrastructure.

Organizations should therefore establish explicit routing policies:

1. Classify the data involved in the task.
2. Prefer a local model for confidential or regulated information.
3. Minimize the context sent to external providers.
4. Grant agents only the MCP tools required for their role.
5. Begin with read only access.
6. Review logs, model usage and agent outcomes.
7. Require separate authorization before enabling mutating operations.

Infinito.Nexus 14.0 supplies the technical boundaries for such a policy. Governance and accountability remain organizational responsibilities.

A Foundation for Sovereign AI Operations

Infinito.Nexus 14.0 connects three layers that companies previously had to assemble themselves:

• A model plane combining local inference and external frontier models
• A tool plane connecting AI to enterprise applications through MCP
• An agent plane deploying Hermes and OpenClaw as isolated virtual employees

Together, these capabilities create a practical path from experimental AI chat to governed AI operations.

Companies no longer have to choose between closed cloud assistants and isolated local models with no access to the systems where work happens. They can keep sensitive workloads on their own infrastructure, selectively use frontier models when appropriate and give agents access to explicitly approved tools.

That is the central promise of Infinito.Nexus 14.0: useful AI where the organization remains in control.

Read the complete Infinito.Nexus 14.0 changelog, explore the source repository, or contact the Infinito.Nexus team to discuss a sovereign AI deployment.