One endpoint in front of everything
Your agent configures a single URL and a single key. Behind it, the gateway resolves which tools that key may see, injects the right credential for each call, and writes the whole thing to an audit log.
Your agents
One URL + one scoped key each
VectorHub Gateway
mcp.vector-hub.org/v1
Scope check
Credential injection
Tool discovery
Audit log
Your systems
Real credentials never leave the vault
Six jobs your agent should not be doing
Each of these is a thing every team rebuilds badly, once per integration, until they stop.
Scope enforcement
Every request is checked against the calling key's connector and tool allowlist before anything else happens. Out-of-scope tools are not merely rejected: they are never visible.
Credential injection
The gateway swaps your bearer token for the real provider credential at call time, refreshing OAuth tokens transparently when they expire.
Tool discovery
The catalog is ranked against the task so the agent receives a working set of tools rather than a wall of schemas.
Response shaping
Large payloads are truncated and summarised at the edge with a handle for full retrieval, so one verbose API response cannot blow the context window.
Retries and idempotency
Transient upstream failures are retried with backoff. Write operations carry idempotency keys so a retry cannot duplicate an action.
Audit logging
Arguments, latency, result status and calling key for every invocation, queryable in the dashboard and exportable to your SIEM.
What happens between the prompt and the API
A tool call travels through five stages. Four of them exist so that the fifth one is safe.
- 1The client sends a tool call with a VectorHub bearer token.
- 2The gateway resolves the key to a tenant and an allowlist.
- 3The tool is checked against that allowlist: unlisted tools never resolve.
- 4The provider credential is decrypted in memory and attached to the outbound request.
- 5The response is shaped, logged and returned; the credential is discarded.
POST /v1/tools/call
Authorization: Bearer vh_live_9f2c…
{
"name": "salesforce.opportunity.search",
"arguments": { "stage": "Negotiation", "limit": 25 }
}
── gateway ─────────────────────────────
key support-bot (tenant: acme)
scope ✓ salesforce.opportunity.*
auth ✓ oauth2 · token refreshed 4m ago
upstream POST /services/data/v61.0/query
latency 187ms (gateway 11ms)
audit logged · trace_8fa21c
────────────────────────────────────────
{ "results": [ … 25 records ], "truncated": false }Built to sit in the hot path
If the gateway is between every agent and every system, it has to behave like infrastructure rather than a side project.
Multi-region edge
Requests terminate at the nearest region; credential stores are regional and never replicated across residency boundaries.
Graceful degradation
A failing connector returns a typed tool error. It does not take down the session or the other 1,499 connectors.
Warm discovery index
The tool index is kept warm per tenant, so the first call of a session is not the slow one.
Transparent rate limits
Upstream limits are surfaced as structured errors with retry hints instead of opaque 429s.
Gateway questions
Is this a hosted MCP server or a proxy in front of mine?
Both are supported. VectorHub hosts managed connectors as first-class MCP tools, and it can also front MCP servers you already run, so your custom server inherits the same scoping, logging and key management.
Which MCP transport does the gateway use?
Streamable HTTP, which is what current clients prefer for remote servers. Server-sent events are supported for older clients, and stdio is available through the CLI bridge for local development.
What happens if an upstream API is down?
The gateway returns a structured tool error rather than a transport failure, so the agent can reason about the outcome and retry or route around it. Upstream status and error rates are visible per connector.
How much latency does the gateway add?
Median added overhead is under 80ms, excluding the upstream call itself. Scope evaluation and credential lookup are in-memory; discovery runs against a warm index.
Point your first agent at it today
Connect an app, mint a key, and add one MCP server to Claude, Codex or Cursor. The whole path takes about four minutes.
No credit card required · Free tier available · Self-host on request