Skip to main content
xpander runs your company’s AI agents and governs every call they make. Teams use Claude Code, Codex and OpenCode with any model, from their desktop, xpander Chat, Slack, Teams, email or the API. Admins set skills, models, budgets and approval rules once in Platform, and every agent works within them.

​
How every call is governed

Each call is checked against the agent’s skills and policies, gets a skill credential for that call only, counts against the organization’s and the agent’s budgets, and is recorded under the person who asked. Where systems allow, the agent acts with that person’s permissions. Actions marked for approval wait for a named person, four hours by default. Calls are paid with xpander credits, your own provider keys or a person’s own Claude plan. See Govern every agent, Audit and debugging and Which key pays.

​
Where to start

​
Start in minutes

​
From your coding tool

Connect Claude Code, Codex, OpenCode or another MCP client, such as Cursor, to https://omni.xpander.ai/mcp. You sign in through OAuth, with no key to paste. On a self-hosted install, your platform team gives you the address. Then, through Omni, you can give tasks to your organization’s agents and read the results, search your organization’s knowledge, or hand a job to the cloud: “Deploy this agent to xpander as a scheduled task.” To send your coding tool’s own model calls through your organization’s models and budgets, create a key in AI models and keys.

​
In xpander Chat

Sign up or sign in at chat.xpander.ai, or at your organization’s own address on a self-hosted install. Pick an agent and give it a task. It runs in an isolated sandbox and keeps running after you close the tab. In the composer, choose the Harness and model. Switching keeps the conversation.
Harness and model picker with Claude Code selected, the providers Google AI Studio, Amazon Bedrock and xpander.ai vllm, and Amazon Bedrock models with Opus 5.5 checked

The Harness and model picker: harnesses on the left, then providers, then that provider's models.

Approvals sets what waits for you: Automatically approve runs every action, Manually approve holds the first shell command, the first file change and, on Claude Code and OpenCode, the first web search or page fetch until you approve, and one approval covers later ones of the same kind in the conversation. Gated commands set on the agent always wait.
Approvals menu open above the composer: Manually approve, checked, with what waits for approval, Automatically approve, and the note Changes this agent for everyone who uses it

The Approvals menu in a conversation's composer: Manually approve or Automatically approve.

Skills are set on the agent, not in the composer. See Create and customize an agent.
Change it in the Harness and model picker. Cursor connects only as a desktop client, not as a harness.

​
Next steps

Understand agents and conversations

How an agent runs your work.

Find your way around xpander Chat

Tasks, approvals, sharing and Omni.

Give an agent skills

What an agent can reach.

Set up and govern xpander

For admins: govern agents and deploy xpander.