Federal Frontier SRE architecture

Our stack — FFO, Warrant, Remuda, local-first inference, posse/SME specialists on a shared Investigator harness (OpenCode Runner or Claude Code runner).

Audience: operators and stakeholders evaluating the Federal Frontier SRE stack against hosted multi-cluster Kubernetes assistants.

This page leads with our architecture. Hosted assistants are a useful reference for operator expectations; they are not the design north star.


Federal Frontier architecture (lead)

Pillar Role
Primary path Problem (English brief or alert) → sre-dispatch Job → Remuda → Warrant/FAS
Dual runners OpenCode Runner or Claude Code runner on the shared Investigator harness
Purpose-built Jobs Task-shaped SRE agent Jobs (Remuda slice + first tools + playbook/scout brief)
Shared harness One harness shape + posse roles — specialization is configuration
SME catalog Automatic routing for known classes — not separate binaries or an agent farm
Posse roles Scout / Sage / Wrangler / Marshal productized onto Remuda tool slices and prompts
Remuda (MCP) Product MCP registry/router (Wanaku engine); Job selects the specialist namespace slice
FFO TypeDB facility graph for topology and remediation scope
Warrant / FAS Statute-backed risk and authz before remediation
Local-first Gemma hot path on-prem; Bedrock Opus cold path with transcript
Estate VitroAI · EKS and RKE2 · Outrider · Ceph · network · platform control — see Coverage

The catalog currently lists five specialist routes on the primary path. See ADR-033 and /docs/sre/sme-catalog/.


Where hosted multi-cluster assistants typically differ

Hosted multi-cluster K8s assistants often optimize for SaaS multi-cluster UX, vendor-hosted correlation, and a large packaged playbook library in public-cloud environments.

Federal Frontier optimizes for sovereignty, facility truth (FFO), Remuda MCP on FMC, Warrant authz, and local-first inference — constraints that define the product shape.

Concern Federal Frontier
Air-gap / sovereignty Local Gemma scout; cloud only on quality-gate miss
AuthZ Warrant PEP + FAS / Keycloak OIDC on Remuda
Facility truth FFO TypeDB, not events alone
Tool plane Remuda namespace slices per specialist Job
Runtime model Shared harness; OpenCode Runner or Claude Code runner; specialist routes accelerate known classes

What we do not claim

  • Drop-in UI parity with any hosted multi-cluster assistant.
  • That every incident class already has a dedicated Remuda tool pack online.
  • That catalog size is a substitute for lab verification.

Grow capability by wiring Remuda tools and classifiers on the shared harness — Jobs stay purpose-built by configuration; runners stay OpenCode Runner or Claude Code runner.


Eupraxia Labs · Federal Frontier · Platform docs