Govern every agent

The controls you set once and every agent inherits.

Admins set these controls once, and every agent inherits them. Screenshots show sample data.

Who may do what

Platform > Organization > Members & groups gives each member a role and puts members in groups with shared defaults.

Role	What they do
Root access	Everything, including organization settings and members
Admin	Manage members, skills, providers, budgets and the settings on this page
Member	Use and assemble agents within what admins allow

Invite members by email. With single sign-on, groups can follow your identity provider’s groups (OIDC group sync), so joiners and leavers are handled in your directory.

Platform > Organization > Members & groups. Shown with sample data.

Sign-in

Deployment	Sign-in options
Cloud and hybrid	Email code, Google, GitHub, or your own identity provider over OIDC (Platform > Organization > Access, an enterprise feature)
Air-gapped	A bootstrap password, email codes, or your OIDC provider, all inside the perimeter

Registration on xpander’s cloud is self-serve and free at chat.xpander.ai. Hybrid and Air-Gapped organizations are set up by their own admins. Setup: SSO with your identity provider.

Platform > Organization > Access > Single sign-on. Shown with sample data.

Skills: what an agent may reach

Admins decide which skills exist and how each connects. Add skill on an agent offers only those. Adding a skill to an agent needs no approval. Admins can set any skill to Off, which removes it from every agent and refuses its calls. Skills served over MCP follow the same rule.

Connection	Whose permissions
Organization-wide	One shared account’s
Personal	Each person’s own, where systems allow

Platform > Agents > Skills registry: what agents and clients can call, each item's type, source and setting. Shown with sample data.

Platform > Agents > Skills registry controls:

Control	What it does
Add item	Adds a library skill, your own API from an OpenAPI spec, a Markdown skill (written, or imported from SKILL.md or a ZIP) or an MCP endpoint
Agent-built skills	Agents build skills off: no agent can build one, existing items stay. Publish without approval off: an agent-shared item waits for an admin. On a self-hosted install, the same two switches are also on Platform > Organization > License, under Organization features, as Agent-built skills and Publish agent-built items without approval
Connect clients	One address for Claude Desktop, ChatGPT or any MCP client, with each person’s own sign-in. On Air-Gapped it needs global.mcp.publicUrl in the install values
Item setting	Auto-injected, On by default, Optional or Off, with the audience that sees it

Agents can build custom functions, Markdown skills and API skills as drafts, test them and publish them here.

Approvals: risky actions wait for a named person

Control	What it holds
On a skill, “These actions need approval”	The listed actions, with who approves
Gated commands	On Claude Code, Codex or OpenCode: any shell command matching a pattern
Permission mode (Approvals in the composer)	Manually approve holds shell commands, file changes, and web fetches and searches on Claude Code and OpenCode. Gated commands wait in both modes

Settings > Approvals: Waiting for you and Decided. Shown with sample data.

With no decision within the hold window (four hours by default, configurable), the action does not run and the run continues. A later decision can still approve a pending gated command. Every decision is recorded with the approver’s name, including self-approvals allowed for unattended agents. See Approvals.

Credentials: the agent never holds a secret

Platform > Agents > Secrets stores every credential encrypted. Agents reference a secret, and the value is injected at the moment of use, never in the prompt, files or record. An external vault (Azure Key Vault today) can be the source. Model provider keys live under Platform > AI governance > Models and follow the same rule.

Platform > Agents > Secrets: names, location and access; values are never shown. Shown with sample data.

Models

Every model call goes through the LLM Gateway. Agents reach only the providers and models you configure in Platform > AI governance > Models, which shows the gateway’s Base URL, each provider’s key status (Connected or Needs a key) and who may use each model. A call to a provider with no key is refused with a message naming it. Every call is logged in Platform > AI governance > Logs.

Platform > AI governance > Models: the Base URL and each provider's key status. Shown on an Air-Gapped install, where the Base URL is http://llm-gw:4000. On a Hybrid install with the release name `xpander`, it is http://xpander-llm-gw:4000. Shown with sample data.

Budgets

Set limits in Platform > AI governance > Budgets for the organization, groups, members, API keys and agents (Budgets). See spend in Platform > AI governance > Usage. A call that would exceed a blocking budget is refused and says so.

Platform > AI governance > Budgets: overview and spending limits by scope. Shown with sample data.

Access to agents and what leaves the workspace

Each agent is Private, a group, or Org-wide. Manage access names who may use it.

Manage access on an agent: private, a group, or the whole organization. Shown with sample data.

The record

Place	What it holds
Every task	Every message, skill call with parameters and result, and model call with cost, under the person who asked. Also over the API: GET /v1/tasks/{id}/thread and /llm_usage
Platform > AI governance > Audit log	Organization events with who did them: agents, skills, vault secrets, users and invitations, groups and OIDC sync, conversation sharing, organization settings. Filterable, exportable as CSV or JSON

Records live in xpander’s cloud, or in your cluster on hybrid and air-gapped installs. See Audit and debugging and Access control.

Change settings by asking Omni

Admins can ask Omni to change most of these settings, from the Omni sidekick or an MCP client. Omni stays within your role, shows each change before and after, applies it only after you confirm, and records it in Audit log under your name.

Where each control lives

Control	Setting
Roles and members	Platform > Organization > Members & groups
Sign-in and identity provider	Platform > Organization > Access. Self-hosted: SSO
Skills your organization made available, including MCP endpoints	Platform > Agents > Skills registry, Add skill on the agent, Settings > Skills > MCP servers, Settings > Connections
Approvals, gated commands, hold time	Skill panel, Agent settings > Advanced > Commands, each approver’s Settings > Approvals
Credentials	Platform > Agents > Secrets, workspace secrets on the agent (Credentials)
Models and provider keys	Platform > AI governance > Models
Budgets and usage	Platform > AI governance > Budgets and Usage
Who may use an agent	ACCESS row and Manage access
Record	Session Manager, Platform > AI governance > Logs and Audit log
Where agents run	Sidebar ENVIRONMENTS switcher, Platform > Agents > Runtime

Next steps

Set budgets and blocks

Spending limits.

Set up single sign-on

Connect your identity provider.