Open Source · MIT Licensed

An open-source control plane
for AI agents.

Xians Agent Control Plane (ACP) runs the layer around your agents: tenants, prompts, conversations, webhooks, human tasks, schedules, logs, and metrics. You bring the agents. ACP runs everything else.

MIT licensed. Self-hosted. Agnostic to the agent runtime.

Who uses it

Two kinds of teams use it.

ACP doesn't care whether you sell agents or run them in-house. The control plane is the same either way.

ISVs · Product vendors

Selling AI agents to your customers.

Embed agents in your SaaS or list them in a marketplace. ACP handles the per-tenant layer underneath.

Enterprises

Running AI agents inside your organization.

Run agents for your employees, or for your own customers. Self-hosted on your own infrastructure.

Capabilities

What it does.

ACP runs around your agents. The agent itself can use any framework, any model, and any data layer.

  • Agent lifecycle

    Register, version, deploy, and retire agents from one place.

  • Multi-tenancy

    Tenant isolation with per-tenant configuration, secrets, and rollout state. One deployment, many customers or teams.

  • Interaction layer

    One interface for conversations, webhooks, and human-in-the-loop tasks. Every channel reaches the agent the same way.

  • Prompts

    Version, review, and roll out prompts. Promote between environments and roll back when behaviour drifts.

  • Observability

    Structured logs, metrics, and traces for every tenant, agent, and conversation.

  • Business processes

    Durable processes and schedulers. Long-running workflows survive restarts and retry safely.

Agent Studio

The control plane, with a UI in the box.

Agent Studio is the default UI for Xians ACP — multi-tenant from day one. Here are some of the controls and operations ACP handles for every tenant, every agent, out of the box.

Architecture

In your Four plane Architecture, Xians ACP is the Control Plane.

Modern agentic systems span four planes. Xians ACP owns the control plane — the other three are yours to shape.

Yours
Plane 1 · Interaction

How users and systems reach the agents

  • Conversations & chat
  • Human-in-the-loop tasks
  • Webhooks & events
  • Agent Studio (default UI)
  • Your web / mobile UI

Surface agents anywhere your users already work. Ships with Agent Studio; bring your own UI on top.

Xians ACP
Plane 2 · Control

Where every agent and tenant is governed

  • Lifecycle
  • Multi-tenancy
  • Interaction routing
  • Prompt management
  • Observability
  • Business processes

MIT licensed. Self-hosted. One place for every agent and every tenant — this is the only plane ACP owns.

Yours
Plane 3 · Runtime

Where agents actually execute

  • Any agent framework
  • Any LLM / model provider
  • MCP tools & servers
  • Your cloud or on-prem

Pick the frameworks, models, and tooling that fit each problem. ACP doesn't decide for you.

Yours
Plane 4 · Data

What the agents read and write

  • Integrations & SaaS APIs
  • Internal services
  • Databases & warehouses
  • Vector stores & data platforms

Your data stays where it is. ACP orchestrates access — it doesn't own the data layer.

Out of the box

Non-functional features. Built in. Day one.

Drop ACP between your web server and your agents. You inherit horizontal scalability, zero-incoming-port security, durable fault tolerance, async orchestration, and full observability — without wiring any of it yourself.

System architecture: a Xians Server (Agent Control Plane) sits between Your Web Server and Your AI Agents. The Web Server connects via the internet to Your Product UI, while Xians Agent Studio offers an optional UI for direct user-agent interaction. The Xians Server connects to internal microservices and exposes a Xians Dev Portal.
Xians runs as a dedicated Agent Microservice. Your Web Server talks to it; agents pull work from it on outbound-only connections.
  • Auto horizontal scaling

    Add or remove agent worker containers at any time. Temporal automatically distributes work — no orchestrator to rewrite, no service registry to maintain.

  • Zero incoming ports

    Agents run in a private subnet with outbound-only pull connections. No inbound surface, no firewall holes, no accidental exposure.

  • Fault tolerance

    Built-in retries, timeouts, and durable state. Workflows survive worker restarts, deploys, and infrastructure hiccups — right where they left off.

  • Decoupled by design

    Agents focus on business logic. Routing, queues, conversation history, secrets, and observability live in the platform — not in your agent code.

  • Fully async by default

    Even synchronous webhooks are queued and processed asynchronously. Stateless workers, configurable timeouts, massive throughput.

  • Multi-tenant isolation

    Tenant-scoped state, secrets, prompts, and conversation history with centralised policy. One deployment, many customers or teams.

  • Durable long-running workflows

    Orchestrate multi-step business processes that span minutes, hours, days, or years — with state managed reliably on Temporal.

  • Audit trail & tracing

    Every workflow execution is logged with distributed tracing. Cost, performance, and behaviour — queryable per agent, per tenant.

Why ACP

Open. Neutral. Self-hosted.

  • MIT licensed

    Read the code, fork it, ship it inside your product. No usage caps. No per-seat fees.

  • Neutral on the agent stack

    Use any agent framework, any model, any data layer. ACP works around your choices.

  • Multi-tenant from the start

    Durable workflows and observability are built in, not bolted on later.

Try it locally.

The docs walk through the architecture and how to run ACP on your machine.