DEPLOYMENT & INTEGRATIONS
Plan how AI traffic reaches Milgram, how the platform runs, and who owns the operational controls around it.
CONNECT THE TRAFFIC PATH
Milgram sits between a supported client workflow and the model provider. Integration begins with routing that workflow through the proxy.
Milgram’s provider handlers include OpenAI Chat Completions and Responses, Anthropic, Google, OpenRouter, and Moonshot. This describes implemented API paths; it is not a partnership claim or a promise that every provider feature is covered.
Client configuration, authentication, streaming, tool-call formats, and modality support can differ. Validate the specific API endpoint and client version you intend to use before expanding the rollout.
Investigation uses a separate MCP connection. An authorized client connects to your deployment’s MCP endpoint, completes the supported OAuth flow, and receives tools constrained by token scopes and user capabilities.
Understand MCP permissionsCUSTOMER-MANAGED DEPLOYMENT
| Path | Best suited to | Operational considerations |
|---|---|---|
| Docker Compose | Single-machine evaluations and compact deployments. | Runs the application services with bundled PostgreSQL and Redis, or uses managed services. Plan persistent storage, backup, upgrades, and host monitoring. |
| Kubernetes / Helm | Teams with an existing Kubernetes operating model. | Uses signed, digest-pinned releases and runtime Secrets. CPU is the default; GPU workers require the appropriate NVIDIA infrastructure. |
| Air-gapped delivery | Environments requiring offline installation and controlled artifact transfer. | Signed bundles include runtime images and model artifacts. Verify signatures and checksums before import, and plan an update process. |
Access is currently invite-only beta. Confirm the available release, integration scope, and delivery arrangements with Milgram before selecting a deployment.
A CONTROLLED ONBOARDING SEQUENCE
Identify AI clients, model endpoints, authentication, data classes, and tool workflows. Assign an owner for policy and an owner for operations.
Test representative requests, streaming responses, tool loops, error responses, and timeouts in an isolated evaluation environment.
Confirm session completeness, review detections, and measure false positives before enabling policies that change or stop traffic.
Enable selected masking, blocking, or compression mechanisms. Check business-task outcomes and practice rollback before increasing traffic.
Size the database, worker, ingress, and storage against representative traffic. The current Helm design uses one stateful worker replica with read-write-once volumes for compiled engines and model promotion.
Backend and dashboard scaling should be evaluated alongside shared dependencies. Establish throughput and recovery evidence for your topology rather than assuming a generic availability or latency figure.
Customer releases use immutable image digests. Release procedures include signed artifacts, vulnerability scanning, and software bills of materials.
Agree on backup and restore procedures, secrets management, ownership of release verification, and rollback of an application or detection update. Request delivery-specific evidence during technical evaluation.
Milgram currently offers invite-only beta access. Contact the team to discuss a hosted evaluation or customer-managed deployment and confirm availability, operating boundaries, and commercial terms.
No universal compatibility claim is made. Many provider-shaped workflows can use a proxy configuration, but managed applications, subscription clients, endpoint versions, and modalities can impose constraints. Bring the exact workflow to evaluation.
INVITE-ONLY BETA
Bring your clients, provider endpoints, and operating requirements to a technical evaluation.
