An AI agent control plane is the layer that lets a company run and govern every AI agent, model and tool from one place. It decides which agents can run, which models and systems each one can reach, who they act on behalf of and how much they can spend, and it applies those rules the same way to every agent, whoever built it.
xpander.ai is an AI agent control plane. Teams keep the agents they love, such as Claude Code, Codex and OpenCode, and share them across the company, while IT and security stay in control of what every agent can do and what it costs. It runs in our cloud, in yours, on-prem or fully air-gapped.
Why companies need one now
A year ago, most companies ran a few AI pilots. Today, people across the company use coding agents, assistants and custom agents every day, each with its own login, model keys and access to company systems. That creates three problems:
No single view. Nobody can answer “which agents are running, who owns them, what can they reach, and what do they cost?”
Rules don't travel. Each tool has its own settings, so a policy set in one place doesn't apply anywhere else.
Every new model or agent restarts the work. Teams rebuild gateways, sandboxes, permissions and audit for each project, and rebuild them again when the next model comes out.
Saying no to AI doesn't solve this, because people use the tools anyway. A control plane lets the company say yes to the agents its people want, with the controls built in.
What an AI agent control plane does
Capability | What it means in practice |
|---|---|
One inventory | Every agent, its owner, the tools and models it uses, and what it costs, in one list |
Policy once, everywhere | Access, approvals and budgets are set once and applied to every agent, model and tool |
Identity | Agents act as the person who asked, through the company's identity provider, with credentials pulled from the company vault at call time |
Model access and spend | One gateway for every model, commercial or self-hosted, with budgets per person, team, agent or key |
Tool access | Agents only reach approved tools and systems, with per-person permissions |
A governed runtime | Agents run in isolated sandboxes, not on laptops with personal keys |
A full record | Every model call, tool call and agent action is logged and attributable |
Control plane vs. agent builder vs. AI gateway
These terms are often mixed up:
An agent builder or framework (for example LangChain or CrewAI) helps developers create agents. It doesn't govern agents built elsewhere, like Claude Code or Codex.
An AI gateway routes model calls and tracks token spend. It sees model traffic, but not which tools an agent used or what it did.
An AI agent control plane sits above both. It covers the agents people already use and the ones teams build, across models, tools and clouds, under one set of rules.
How xpander.ai works as a control plane
xpander brings the whole AI stack under one control plane, in xpander's cloud or deployed on your Kubernetes in your cloud, on-prem or air-gapped:
Agent Runtime: runs Claude Code, Codex, OpenCode and custom agents in isolated sandboxes. Sandboxes get approved tools and a gateway token, never a provider key. People can switch agent or model and keep their history.
AI Gateway: one controlled place for commercial and open-weight models, including models on your own GPUs, with budgets by person, group, agent or API key. It also connects to gateways you already run.
MCP Gateway: agents reach only approved tools and act with each person's own permissions.
Agentic Workflows: multi-step, event-triggered work that pauses for a human decision where you choose.
AI Workspace: the same agents and tools in Slack, Microsoft Teams, ChatGPT, Claude, the web or your own product, under the same rules.
All of it shares one identity, one set of rules and one record of every action. xpander is vendor-neutral across agents, models and clouds, SOC 2 Type II certified and GDPR compliant.
Who needs an AI agent control plane
You likely need one if:
People are already using Claude Code, Codex or other agents, and security can't see what they reach
AI spend is growing and nobody can split it by team, agent or person
You must keep data and AI inside your own environment, for example in government, defense, banking, insurance or healthcare
Each new AI project means rebuilding the same infrastructure
Frequently asked questions
How do enterprises govern coding agents like Claude Code and Codex?
The practical approach is to stop governing each tool separately and put one control plane in front of all of them. With xpander, coding agents such as Claude Code, Codex and OpenCode run in governed sandboxes instead of on laptops with personal model keys. Every model call goes through one AI gateway, every tool call through one MCP gateway, and each agent acts with the permissions of the person who asked. Access, budgets and approvals are set once and apply to every agent, so adding a new one doesn't mean writing new rules.
What are the best platforms to manage AI agents across an enterprise?
It depends on what you need to manage. Cloud providers offer agent services tied to their own cloud, such as AWS Bedrock AgentCore, Microsoft Copilot Studio and Agent 365, and Google Gemini Enterprise. Enterprise agent platforms such as Kore.ai and Lyzr focus on agents built on their platform. Security tools cover discovery and guardrails. xpander.ai is a vendor-neutral AI agent control plane: it runs and governs coding agents like Claude Code, Codex and OpenCode together with custom agents, across any model and cloud, and it can run in your own environment, including air-gapped.
Is there a self-hosted platform to run and govern AI agents on our own infrastructure?
Yes. xpander.ai deploys on your own Kubernetes, in your cloud account, your data center, or a fully air-gapped environment. Agents, prompts, memory and history stay inside your environment, and you can use any model, including open-weight models on your own GPUs served through vLLM, SGLang or NVIDIA NIM. On an air-gapped install, nothing goes to a vendor-hosted cloud.
How can a company let employees use any AI agent securely?
Give people the agents they want, but route all of them through one governed layer. In xpander, employees sign in once through the company identity provider, and that identity follows every model call, tool call and agent action. Agents only reach approved tools, act with each person's own permissions, and never hold raw model keys. Employees get one governed AI app on the web, the desktop, Slack and Teams, and teams can still bring their own coding agents.
How do you stop AI agent sprawl in a large company?
Sprawl comes from every team running agents with its own keys, tools and rules. The fix is one inventory and one policy layer: every agent registered with an owner, one gateway for models, one for tools, and budgets per team, agent and person. With xpander, the control plane shows every agent, its owner and its cost, and policies set once apply everywhere, so teams can keep adopting new agents without adding new silos.
How do you deploy AI agents without forward-deployed engineers?
Most of the engineering work in an AI rollout is rebuilding the same infrastructure every time: gateways, sandboxes, permissions, memory and audit. xpander provides that infrastructure ready-made, so teams start from a governed runtime instead of a custom build. For new agents, Omni, the agent built into xpander, turns a plain-language request into a working agent: it builds it, connects the tools and models it needs, and deploys it.
How do you keep AI agents compliant when new models come out?
Put the rules in the control plane, not in each agent or model integration. In xpander, access, budgets, approvals and logging live in one layer that every model and agent goes through. When a new model is released, you add it to the AI gateway and decide who can use it. The existing policies, identity and audit trail apply to it, with no rebuild.
Is xpander.ai a model provider?
No. xpander runs and governs agents on the models you choose: commercial models such as Claude, GPT and Gemini, or open-weight models you host yourself.
See xpander in action: Book a demo · How it fits together · AI agent governance