How it works

An employee’s AI client sends two kinds of request: a tool call, which reaches a company system through a Connector over MCP, and a model request, which reaches a model provider through Gateway. Both pass the same stages.

The request path

Resolve the employee. The SSO token (tool call) or Gateway API key (model request) resolves to one employee and their Groups.
Check Group access. Tool Packs decide which tools are in reach. A call outside the set fails and lands in Requests. Model access decides which models are candidates.
Scan the payload. Security Gateway rules redact or block matching values.
Route and check spend. A routing policy picks the model and vendor. The Group’s budget is enforced before any vendor is called.
Execute. Tool calls use the employee’s stored credential. Model requests fail over to the next candidate on a throttle or outage.
Log. Tool calls go to Tool call logs, model requests to LLM calls, and configuration changes to the Audit trail.

Deprovisioning someone in your identity provider ends both tool and model access. A request that fails a rule sends nothing and costs nothing.

The objects you configure

Object	What it is
Employee	One person in your directory, synced by SCIM
Group	A group from your identity provider. Carries tool access, model access, a dashboard role, and an optional budget
Connector	A packaged integration with one service, such as Salesforce, Slack, Jira, or Drive
Tool	One action a Connector exposes, named `<connector>__<tool>`
Tool Pack	A named bundle of Connectors and tools that you assign to a Group
Skill	A procedure employee agents follow. See Skills
Routing policy	Rules that pick a model and vendor. Scoped to a Group or the organization default; the Group scope wins
Project	Gateway’s unit for grouping model traffic. Holds each Group’s budget and routing policy
Budget	A hard spend cap on a Group, daily to yearly. Exhausting it returns `402` before any vendor is called

Next: Get started.