You can build all of this yourself. Here’s the parts list.
Every piece of Omnia has an open-source counterpart, and a platform team can assemble them. What takes the time is the glue between the parts, which you write yourself and maintain for as long as the agents run.
Four pieces of glue that decide whether it works
One version for prompts, tools, policy and tests together, carried into every trace, eval result and audit record.
Feed the eval suite’s result and live agent metrics into promotion, and keep each conversation on the version it started with.
Carry the caller’s identity through the agent to every tool, MCP server and remote agent it calls.
Make a user’s consent withdrawal reach the memories already written, and make newer facts win over old ones.
The parts list
Untick what you don’t need. The total underneath keeps count.
| Include | You need | Open-source part | Glue you write | In Omnia |
|---|---|---|---|---|
| LangGraph compare | Serve it over the protocols your callers use, drain conversations when a pod goes away, and scale on load. | Five ways to serve an agent, Graceful drain, Autoscaling docs | ||
| hard part | Git, an OCI registry and Argo CD | One version for prompts, tools, policy and tests together, carried into every trace, eval result and audit record. | GitOps content sync docs | |
| LiteLLM compare | Budgets per agent and workspace, and provider credentials from workload identity. | Seven native API interfaces, Any OpenAI-compatible endpoint, Enforced cost budgets, Platform auth docs | ||
| OpenTelemetry Collector and Langfuse compare | A span inventory that means the same thing in every agent, and dashboards for it. | Distributed tracing, Metrics and dashboards, Live topology docs | ||
| Promptfoo compare | Point it at your real tools in a test environment, on models you’ve approved, with assertions on what the agent did. | Simulated users, Adversarial personas, Tool-call assertions docs | ||
| Hamming compare | A second vendor, or a SIP harness of your own that interrupts, talks over and goes silent. | Hostile callers on a real voice line docs | ||
| hard part | Argo Rollouts | Feed the eval suite’s result and live agent metrics into promotion, and keep each conversation on the version it started with. | Deploy-gate wiring, Progressive rollouts, Sticky sessions, Metric-gated promotion docs enterprise | |
| Open Policy Agent compare | Put a decision point in front of every tool call, with the arguments and the end user in the input, and fail closed when it can’t answer. | Per-call tool policy, Claim-based access docs enterprise | ||
| hard part | Keycloak and oauth2-proxy | Carry the caller’s identity through the agent to every tool, MCP server and remote agent it calls. | OIDC auth, Edge-trust auth, End-user identity propagation, Outgoing auth docs | |
| Microsoft Presidio | Run it before the session is recorded, before memory is written and before the model sees a tool response. | PII redaction in sessions, PII redaction on memory writes, Tool response redaction docs enterprise | ||
| hard part | Mem0 compare | Make a user’s consent withdrawal reach the memories already written, and make newer facts win over old ones. | Cross-session memory, Tiered recall, Consent-aware memory docs | |
| Your SIEM and a runbook | A record of what happened on whose authority, and an erasure that reaches sessions, memory and archives. | Audit logging, DSAR erasure, Privacy audit hub docs enterprise | ||
| Namespaces, RBAC and NetworkPolicy | Per-team isolation, budgets and grants that the agent layer understands. | Workspaces, Network policy docs |
Keep the parts you like.
Omnia emits standard OpenTelemetry, reaches any OpenAI-compatible gateway and runs agents built in other frameworks, so a stack you’ve already started isn’t wasted.
Parts checked 2026-10-22. Each links to the project’s own site.