User guide

xpander is the governed runtime for your organization’s AI agents. Admins approve skills and models, set budgets and approvals, hold credentials and keep the record in one place. Agents run inside those decisions: every model call and every skill call passes through them as it happens. You reach it in one of two ways.

​

1. Launch a cloud session

Open Xpander Chat, pick an agent, and give it a task. On xpander cloud it lives at chat.xpander.ai: your organization’s admin invites you, and a new organization is activated by the xpander team after a short call booked from the sign-in page. A self-hosted install serves it on your organization’s own address. xpander launches a session of Claude Code, Codex or OpenCode for that run in its own isolated sandbox, with your organization’s policies applied: the approved skills, the models and budget set for you, approvals on risky actions, and the record. You shape the session in the composer:

• Harness and model: the picker lists the providers your organization enabled and every model with the harness it runs on. Switching keeps the conversation; the new harness receives the history.

  

• Permission mode: Full runs every command the agent decides on; Gated holds shell commands, file edits and web fetches for approval. Gated commands your admins set always wait.

  

Which skills the agent may use is set in its Agent settings, not the composer: Add skill there offers only what your organization approved. See Create and customize an agent. The session runs unattended after you close the tab. Its task shows every action with status and time, the files it produced, token usage and cost, and the result. Teammates can open the same conversation and continue it. The same sessions start from Slack, Teams, email, a webhook, a schedule or POST /v1/agents/{agent_id}/invoke. See Xpander Chat, Tasks and Approvals.

​

2. Use your desktop harness

Keep working in Claude Code, Codex, Cursor or OpenCode on your laptop. Connect the client to xpander once, at https://omni.xpander.ai/mcp (OAuth in the browser, no key to paste). On a self-hosted install your platform team gives you the address once the install exposes an MCP origin. The endpoint is served by Omni, the built-in agent, so from your client you can:

• Task the organization’s agents and read their results, with the skills and permissions your admins set for them.
• Search the organization’s knowledge, and have an agent assembled from the skills your admins approved.
• Hand a job to the cloud. When a job should not depend on your laptop, ask in plain words: “Deploy this agent to xpander as a scheduled task.” xpander creates the agent, imports the skills it needs, starts it as a cloud session and tells you where it lives. From then on it runs and answers in Xpander Chat, Slack, Teams or email.

Per-client setup is on Connect your MCP client. Where systems allow, an agent acts with your own permissions in the outside system, and every action is recorded under your name. The client’s own model calls and local skills stay on the laptop today; governing them there is on the roadmap, described on Desktop AI clients.

​

Which harness a session gets

A new agent starts on Claude Code, Codex or OpenCode, following your organization’s default model: an Anthropic default gives Claude Code, an OpenAI default gives Codex, any other provider gives OpenCode on that model. The composer’s Harness and model picker shows the choice and lets you change it. Cursor is not selectable as a harness yet; it connects as a desktop client.

​

Next

• Agents and conversations: what an agent is, what launching one does, and what is isolated.
• Xpander Chat: what is on the screen.
• Skills: what an agent can reach, and how to connect your own account where systems allow.
• Approvals: what waits, who decides, what happens meanwhile.