Skip to main content

Production quickstart

This walkthrough gives an org owner or org admin a repeatable first production workflow: one agent, one toolkit, a safe read, an approval-gated write, and an audit check.

1. Prepare the organisation​

Open CortexHub Cloud and confirm that you are an owner or admin. Invite the people who will operate the workflow and put them in a dedicated group. The role definitions explain which changes each role can make.

2. Enable and publish one toolkit​

In Org space -> Toolkits, enable one catalogue toolkit or create an organisation custom OpenAPI toolkit. Then publish it to the group that will use it. Configure the allowed auth method, OAuth scopes, and tool policies before publishing. Catalogue publication alone does not grant access.

3. Create the agent​

Choose the runtime that matches your workflow:

  • Hosted Org Agent: create it under Org space -> Agents. CortexHub runs it and it can be installed as a named Slack bot.
  • SDK Org Agent: register it under Org space -> Agents, then run your process with an organisation gateway key bound to that agent.
  • Member-authenticated runtime: use a programmatic Member Credential when the process should act as one CortexHub member.

Assign the Org Agent to its own agent group. Publish the toolkit to that group and configure an organisation-shared credential when the toolkit needs one. Org Agents do not borrow a member's personal connection.

4. Install Slack when the agent is hosted​

For a hosted Org Agent, open Settings -> Integrations -> Slack, connect Slack once for the organisation, then open the agent's Slack tab and choose Add to Slack. Authorize the generated named app and add it to a test channel. Bot installation does not grant toolkit access; the agent group and policy still govern each call.

5. Run the read test​

Ask the agent for a small, reversible read such as listing a project or finding a recent message. Confirm that:

  • the tool is visible only because the agent group is entitled to it;
  • the call result is ok;
  • the agent activity view shows the tool, principal, policy decision, and receipt.

6. Verify approval and denial​

Try one write action with a harmless test record. Configure that safety profile as require approval and destructive actions as deny. Confirm that the write parks as an MCP Task, the configured operator or user approves it, and the task completes once. Then try a destructive action in a safe test context and confirm that policy denies it before the provider is called.

7. Connect an external runtime (optional)​

For a backend with its own users, use the organisation gateway key and send X-CortexHub-End-User-Subject on every request. For a backend acting as one CortexHub member, use a programmatic Member Credential. Interactive Claude, ChatGPT, and Cursor connections use browser OAuth with Authorization Code + PKCE.

8. Check the operating posture​

Before adding more tools or channels, review the agent's sessions, approvals, audit records, and usage. Configure an approval channel or OTLP/OCSF export if your operators need independent notifications or retention. Use Operational limits and reliability to set retry, monitoring, and incident thresholds.

See also​