Direct authorized route
API Management passes an authenticated request straight to a warm CGP runtime, skipping the queue whenever the caller needs an answer right away.
Technology and architecture
CGP Gateway implements the Connectivity Gateway Protocol (CGP) as a single-region Azure private beta: a multitenant control plane with synchronous and durable asynchronous routes; signed transactions with receipts and tenant-aware metadata; managed identities; and full operational telemetry. Inside the gateway, the runtime converts connection-oriented TCP traffic into encrypted UDP datagrams (DTLS 1.3) built for machine-speed exchange.
Architecture narrative
The beta runs on managed Azure services and avoids orchestration the product does not yet need.
The proprietary CGP runtime is built on HyperCycle technology and sits behind the public API boundary. This page describes each service's role and how data moves between them. It does not describe the protocol's internal mechanics, message formats, or algorithms.
Azure API Management protects, meters, and observes the external API surface. The proprietary CGP runtime does the connection work behind that boundary.
Inside the runtime
Every transaction runs through the same seven-stage pipeline, from protocol adaptation to receipt and settlement reference.
Two execution paths
The architecture separates fast synchronous requests from asynchronous work that must survive retries, downstream delays, or fan-out.
API Management passes an authenticated request straight to a warm CGP runtime, skipping the queue whenever the caller needs an answer right away.
Service Bus stores work until a connector job can process it, supporting retries, scheduled work, and event fan-out.
Microsoft Foundry or Azure OpenAI may be an authorized destination, but the gateway is not limited to one AI provider or model family.
Payload content is processed in memory and not persisted by default; required audit metadata is retained under a stated schedule.
Beyond the boundary
The gateway routes work only to destinations a workspace authorizes, and today that means HTTPS endpoints on existing rails. Because the runtime is built on HyperCycle technology, the same gateway is designed to reach native execution on the Internet of AI when measured economics justify it. The gateway is an access layer above that network, not a node inside it.
The substrate capabilities below are design intent, not measured fact.
Registered HTTPS destinations, agent frameworks, and model services receive routed work today, with settlement references recorded in the transaction and value staying at the edge.
The runtime is built on HyperCycle technology. Boundary nodes are designed to admit external traffic into the machine network, and network nodes to execute and settle native work.
Substrate protocols such as TODA/IP are designed to settle value peer to peer with local finality, keeping chain finality out of the inner loop of high-frequency work.
Repeated, high-frequency work becomes a candidate for native execution only where measured economics and customer requirements justify the move.
Current versus planned
Active services run today’s private beta. Each planned service is tied to a specific scale or customer requirement.
Public TLS termination, global entry point, health-aware routing, and WAF policies.
ActiveAuthentication policy, request limits, versioned API surface, metering, and API telemetry.
ActiveCGP control API and warm connection runtime with revisions and bounded autoscaling.
ActiveDurable asynchronous jobs, dead-letter handling, retries, and event distribution.
ActiveTenant-partitioned endpoint, policy, route, transaction metadata, and receipt records.
ActiveEncrypted audit exports, test artifacts, and lifecycle-managed backup packages.
ActiveApplication Insights and Log Analytics for latency, failures, queue depth, alerts, and operating reports.
ActiveWarm recovery environment, replicated configuration, and rehearsed regional failover.
PlannedDedicated premium customer accounts for stronger data and throughput isolation.
PlannedData lifecycle
The private beta stores configuration, transaction metadata, and receipts. By default, it does not store customer payload content.
| Data category | Purpose | Retention |
|---|---|---|
| Identity and tenant records | Authenticate users, assign roles, and resolve the correct workspace. | Account life plus 30 days. |
| Endpoint and policy configuration | Define allowed systems, routes, limits, and connection behavior. | Until deleted plus 30-day recovery window. |
| Transaction metadata and receipts | Captures the transaction ID, route, timing, result class, policy decision, evidence pointers, and receipt reference for each transaction. | 30 days standard; configurable to 7 days. |
| Customer payload content | Transient routing only. | Not persisted by default. |
| Audit exports | Customer-requested evidence packages. | 90 days, then lifecycle deletion. |
Performance method
Gateway overhead is measured separately from downstream model and agent latency.
Median gateway overhead.
p95 gateway overhead.
p99 gateway overhead.
Successful routing rate.
Methodology: 450,000 requests over 30 minutes at 250 requests per second; 4 KB request payload; one Azure region; four warm Container Apps replicas; fixed test destination; authentication and policy evaluation included; downstream model latency excluded. These are internal measurements taken under the method above. An independent benchmark is planned; until then we publish no comparisons with other systems. Full detail is on the Results page.
Trust layer
The Trust page walks through identity, isolation, encryption, secrets, monitoring, backup, incident response, and AI-use boundaries as they stand today in the private beta.
Review security controls