Questions, answered properly

The Model Context Protocol is an open standard for exposing tools to AI agents. Normally each service needs its own MCP server, its own credentials and its own client configuration. VectorHub runs one server that fronts every connector you enable, so an agent configures a single endpoint and gains access to everything behind it.

No. That is the point of the gateway. You authenticate each service once in VectorHub, then mint a scoped key per agent. Claude, Codex, Cursor and your own runtimes each get a key that grants only the connectors and tools you chose. Provider secrets never leave the vault.

They do not, which is why VectorHub does not send them all. Tool discovery ranks the catalog against the current task and exposes a working set, so the agent sees the tools it needs rather than every tool that exists.

Yes. Point the API → MCP converter at an OpenAPI or GraphQL schema, or describe the endpoints manually, and VectorHub generates a typed MCP server with auth attached. Private connectors stay scoped to your organisation.

Access stops immediately, everywhere that key was configured. Because agents hold a VectorHub key rather than provider credentials, revocation is a single action instead of a rotation project across every service.

OAuth 2.0 with PKCE and refresh rotation, API keys and bearer tokens, username and password for basic-auth services, and custom schemes such as HMAC request signing or mTLS.

Yes. Enterprise deployments can run in your VPC or fully self-hosted, with data residency controls and audit export. Talk to us about the deployment model you need.

MCP over streamable HTTP is the primary interface and works with any compliant client. A REST API and TypeScript and Python SDKs cover provisioning, key management and log access.

In an encrypted vault, using envelope encryption with a per-tenant data key wrapped by a managed KMS root key. Plaintext exists only in memory for the duration of one outbound call, and credential material is excluded at the serialiser so it never reaches a log.

It becomes bounded. Nothing reliably stops an agent from being steered by untrusted input, but a call outside the key's allowlist is rejected before any credential is touched, and the attempt is logged against a named key.

Yes. Enterprise deployments run in your VPC or fully air-gapped, with connector packages delivered as signed bundles. Credentials never leave your boundary.

By tool calls through the gateway. Connectors you have enabled but are not calling cost nothing, and denied calls are never billed.

No. Three connected apps and one scoped key stay free indefinitely: enough to build something real on a side project.

Yes. Early-stage companies and academic research groups get significantly discounted terms; mention it when you get in touch.