Architecture & security
CortexHub is a governed control plane and MCP gateway for AI agents. Your AI client, CortexHub Co-worker, Org Agent, or application runtime sends tool requests to the gateway. CortexHub resolves identity, access, policy, and approval before a provider call is made.
Request path
AI client, Co-worker, Org Agent, or application runtime
↓
MCP Gateway
↓
Authentication and agent identity
↓
Organisation and group entitlement
↓
Policy and safety evaluation
↓
Approval or step-up when required
↓
Credential injection at invocation time
↓
Provider API
↓
Audit event, enforcement receipt, and optional telemetry export
The same governance path applies whether the agent is hosted by CortexHub or runs in your own application. A denied or approval-gated request does not reach the provider until the required decision is recorded.
Runtime choices
| Runtime | Identity and access |
|---|---|
| Hosted Co-worker | Belongs to one member and uses that member's governed connections. |
| Hosted Org Agent | Belongs to the organisation, uses an Org Agent group, and can be installed as a named Slack bot. |
| SDK Org Agent | Belongs to the organisation and runs in your application with an organisation gateway key. |
| External MCP client | Uses interactive Member OAuth or a programmatic Member Credential, depending on whether a person is present. |
| External-user backend | Uses an organisation gateway key and attests the external user on every call. |
Security boundaries
- Every request is evaluated in an organisation context. Group entitlements determine which toolkits and tools a principal may reach; layered policies then resolve permit, approval, or deny.
- Provider credentials are stored securely and injected only while a governed tool runs. Agents and chat transcripts do not receive raw vendor secrets.
- Interactive MCP OAuth, programmatic Member Credentials, organisation gateway keys, and Platform API OAuth are separate credential models. See Choose your authentication model.
- Policy decisions, approvals, denials, redaction outcomes, and tool results produce audit and enforcement records. Optional telemetry destinations can deliver spans to your own collector or SIEM.
Enterprise review
Use the product controls and audit evidence described here to evaluate CortexHub against your organisation's security requirements. Contract-specific commitments and security review materials are available from CortexHub during evaluation.