Skip to main content
Connect Claude Code, Claude Desktop, ChatGPT, Cursor or any MCP client to Omni, the agent built into every organization, and work with it without leaving the client. Every call runs with your permissions, where systems allow.

Connect from a client

Sign-in is OAuth 2.1 only, no API key. Most clients, including ChatGPT, Claude Code, Claude Desktop and Cursor, find the sign-in from the URL and open the browser. You sign in once. The client keeps the token.
Add the endpoint once from the terminal, then sign in from inside a session:
In Claude Code, run /mcp, pick omni and complete the browser sign-in. From then on, ask in plain words: “Ask Omni what my agents did today” or “Deploy this agent to xpander as a scheduled task.” On a self-hosted install, replace the address with your installation’s MCP origin.
This is not the API’s MCP endpoint (mcp.xpander.ai, also OAuth 2.1), which exposes the xpander API itself and is not an agent. The Omni endpoint is scoped to you: your conversations and your access. For a direct connection to one agent, use the API’s MCP endpoint.

Calls

Pass the conversation_id back to ask_omni (or ask_omni_async) to continue the same thread. Runs can take minutes: if one may exceed your client’s timeout, start it with ask_omni_async and poll get_omni_task until it’s done. Everything else, such as finding and tasking your agents, building or editing agents, managing tasks and schedules, and searching your organization’s knowledge, you ask Omni for in plain words. Omni runs as you, so your own access bounds what a request can touch.
MCP clients show plain Markdown, so Omni answers in self-contained Markdown and never raises an interactive card. If it needs something it can’t get here, such as a missing credential, it does what it can and tells you how to fix it in xpander Chat. If a skill needs authorization mid-run, ask_omni returns the link with the conversation_id. Authorize, then fetch the result with get_omni_task. These runs show in Session Manager tagged by client, for example “MCP - ChatGPT” or “MCP - Claude”.

Admin actions

Admins can read and change organization settings over the same endpoint, the same actions as the Omni sidekick. Your client lists only the calls your Platform role covers, and each is checked again when it runs.
Each Platform area is one call that takes an action:
Reads answer at once. A change takes two calls:
  1. The first returns a preview (now and after) and a confirm_token. Nothing changes yet.
  2. Call again with that confirm_token and confirm set to true to apply it, or false to drop it.
The token works once, only for you and that call, and expires after about 10 minutes.Secret values and new API key values never pass through. Applied changes go to the Audit log under your name. A connection with a bare API key has no person behind it, so admin calls refuse it.

Troubleshooting

The endpoint is OAuth-only and rejects x-api-key or any static key. Remove key-based config, re-add the endpoint and sign in in the browser.
Check the URL is exactly https://omni.xpander.ai/mcp and finish sign-in in the window the client opens. If a token is stale, remove and re-add the endpoint.
Open the link in Omni’s reply and authorize in xpander Chat. Then call get_omni_task with the same conversation_id.

Next steps

Call your agents over the MCP API

The API’s MCP endpoint.

Learn what Omni can do

Omni across xpander.