One gateway.Full control.

Schedule demo
Contact via email ›

A gateway for models and MCP tools, on infrastructure you own.

Cursor, Claude Code, and your own apps call one URL. Keys, audit, and PII stay with you.

learn more

Ready to see
it for yourself?

Schedule demo
In practice

Measured performance.A named stack.

The hot path is Rust. Permission is checked on every tools/call.

PerformanceOn a fast mock MCP, 3-replica HA holds ~4.4k RPS with zero fail.

  • 4.4k RPS, 3-replica HA, zero fail
  • 5.3k RPS at saturation
  • 0 fails on 21,123 verified calls
Schedule demo
OK RPS503 fail
45.3619.9620.31056.2
c1c30 / 8c30 / 64c100

StackPermission in SpiceDB. TLS at Caddy. The hot path is Rust.

  • Rust on the MCP proxy hot path
  • SpiceDB on every tools/call
  • Kubernetes under the 3-replica HA
Schedule demo
RustSpiceDBCaddyPostgreSQLKubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghornRustSpiceDBCaddyPostgreSQLKubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghornRustSpiceDBCaddyPostgreSQLKubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghorn
KubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghornRustSpiceDBCaddyPostgreSQLKubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghornRustSpiceDBCaddyPostgreSQLKubernetesClickHouseRedisOTelUbuntuHelmcontainerdQdrantLonghornRustSpiceDBCaddyPostgreSQL
FAQ

Frequentlyasked questions

Quick answers on keys, MCP, and what stays on your network.

Point Cursor, Claude Code, Codex, WorkBuddy, or an internal service at one gateway URL. From there they can call the models your company has published, and attach MCP tools — DingTalk, GitHub, Confluence, databases, or your own APIs — without putting vendor keys in the client. The gateway is the hop that holds the upstream credentials, checks permission, and writes the record.

Yes. Clients hold a Gateway Key. Model-provider keys and MCP credentials stay pooled on the gateway and are injected only on the way out. Rotate or revoke a key and it takes effect on the next call. The laptop never had the upstream secret, so there is nothing to chase down when someone leaves.

Admins publish tools into the MCP plaza. A user opens a tool, copies the connection JSON, and pastes it into Cursor, Claude Code, Codex, or WorkBuddy. They can run a tool test on the gateway before the client ever sees it. If a tool’s definition changes in a way that affects safety, it can require a new approval instead of updating silently.

It does not have to. Zentrix installs from an offline package. Bring your own models — vLLM, SGLang, Ollama, or any OpenAI-compatible endpoint — and the platform does not need public outbound access. PII detection runs locally, with allow, audit, redact, or block. Unknown matches fail closed.

A raw endpoint has no company permission model, no quota, no PII fence, and no audit of the person who actually made the call. Zentrix sits in front: one base URL, routing and failover, visibility by workspace, and a log that names the caller. If a client bypasses the gateway, that traffic will not show up in the admin console. That is the point.

Visibility and invocation are separate. A model or tool a user is not allowed to use is not callable. Authorization is checked on every call, down to the individual tool, against the real user — not only the application account. Admins can see who called what, from where, and at what cost. Permission changes take effect without restarting the gateway.