Arvenza Helix Build, run and govern AI agents on the systems you already run.

Helix is the Enterprise AI Operating Layer: agents, integration and knowledge pipelines in one runtime. You describe what should happen in readable configuration. Helix connects it to your applications, grounds it in your documents, and keeps a person in charge of the decisions that matter.

Arrange a Conversation Read the Docs

One self-contained binary. Runs on-premises, in your private cloud, or completely offline with a self-hosted model.

An agent picks up work from any channel, reasons, acts and observes, then updates your systems. Sensitive actions wait for a person.

Agents, integration and knowledge in one runtime

Most enterprise AI is assembled from separate tools for agents, integration, retrieval and monitoring. In Helix they share one runtime, one security model and one audit trail.

Agents

Seven agent patterns: Single agents that reason and act, supervisors that delegate to specialists, planners, deep researchers, and teams that work in sequence, in parallel or in a loop.
Human approval: Any tool can require a person's approval before it runs. Runs are checkpointed, so an agent resumes exactly where it paused.
Memory: Conversation memory that lasts across sessions, from sliding windows and summaries to database and vector memory.
More than text: Agents can work with images, audio and video as well as text.

Knowledge

Documents in, cited answers out: Ingest files, scanned documents, code and web content; chunk, embed and index them; answer with the sources attached.
Retrieval that finds the passage: Hybrid search, re-ranking and query rewriting find the right passage, not just the right document.
Privacy in the pipeline: Personal data can be masked before it is indexed or shown to a model.

Integration

440+ connectors: Across 33 categories: CRM and ERP systems, databases, messaging, cloud services, and any REST or SOAP API.
50+ integration patterns: Routing, transformation, retries, circuit breakers, sagas and dead-letter handling for the moments when systems misbehave.
Starlang: A compact expression language inside the configuration, for conditions, data mapping and policy rules.
16+ model providers: Commercial model services, the major clouds and self-hosted models. Changing provider is a one-line configuration change.

Governance

Tamper-evident audit: Every allowed and blocked action is written to a hash-chained log that an auditor can verify after the fact.
Kill switches and privilege levels: Stop a capability instantly without redeploying, and give each agent only the access it needs.
Policy and prompt defence: Rules decide what an agent may do, a scanner checks the MCP servers it connects to, and a classifier catches prompt attacks.
Enterprise security: JWT and OAuth2, API keys, TLS and mutual TLS, rate limits, IP allow-lists, managed secrets and multi-tenancy.

Open standards

MCP and A2A: Publish any agent, pipeline or workflow as MCP tools, use tools from other MCP servers, and let agents work with other agents over A2A.
AG-UI and A2UI: Stream conversations to web front ends and render cards, buttons and forms, alongside plain REST.

Operations

Visibility: Metrics, OpenTelemetry tracing, dashboards, and latency targets for every workflow and agent.
Always on: Hot reload, graceful shutdown that finishes work in flight, and clustering with leader election.

An agent is a short file you can read

This agent answers on an HTTP endpoint and searches your knowledge base on its own, but it waits for a person before it issues a refund. The last block publishes it as MCP tools, so other assistants and agents can call it.

Run it with helix run. The same file runs on a laptop, a server or a cluster, so what works in the demo is what runs in production.

Configuration lives in version control and is reviewed like code.
Starlang covers the conditions and data mapping that plain configuration can't.
Custom code is only needed for components you build yourself.

Read the getting-started guide to run your first agent, or browse the component reference.

service-desk.yaml

agents:
  - id: service-desk
    type: react
    trigger: http-server://0.0.0.0:8080/agent
    model:
      provider: your-provider   # cloud or self-hosted
      model: your-model
    approval_policy: tagged
    tools:
      - name: search
        type: function
        description: "Search the knowledge base"
        parameters:
          type: object
          properties:
            query: { type: string }
          required: ["query"]
      - name: issue_refund
        type: function
        description: "Refund a customer order"
        requires_approval: true   # waits for a person
        parameters:
          type: object
          properties:
            order_id: { type: string }
          required: ["order_id"]

mcps:
  - id: service-desk-tools
    addr: "0.0.0.0:3100"
    expose: [service-desk]

It runs where your data is

One binary

Helix ships as a single self-contained program with nothing else to install. Run it in the foreground while you build, or as a background service.

Containers and clusters

Package it as a container image and run several replicas with health checks. Nodes find each other, elect a leader and carry on when one of them fails.

Inside your network

Everything runs in your environment. Pair it with a self-hosted model and no data leaves your network; even the licence check works offline.

Works with the rest of Arvenza

AgentManagerPoint Helix's model settings at the gateway to add routing, failover, budgets and guardrails to every model call. Arvenza DBGive agents retrieval by meaning, by exact words and by connection, as MCP tools. VGLLet agents ask graph questions of your ERP and CRM data where it lives, as MCP tools.

See Helix on your own use case

A focused proof of concept in your environment, with your data, says more than any page.

Arrange a Conversation