All sensing, analysis and action is decoupled into 42 Model Context Protocol servers over a three-host hybrid fabric, forming a typed topology graph that is itself a security asset — topology drift is an alert.
| Host | Platform | Role |
|---|---|---|
| laptop | Linux x86_64 · 12 CPU · 30.7 GB | Primary ingestion, SSH jump, agent venv host |
| m1 | Darwin arm64 · 8 CPU · 16 GB | Local AI inference (air-gapped path for PII / hardware telemetry) |
| omen | Linux x86_64 · 24 CPU · 31.1 GB | MCP server farm, heavy simulation, ClickHouse |
Edge layer: Workers + Caddy + tunnels across 49 hostnames / 4 origins; 1,652 discovered services are continuously inventoried.
Every sensor, datastore, tool and orchestrator is an MCP server with typed tools — swappable models, auditable calls, a topology graph queryable by agents (full inventory: Appendix A).
Edge kinds in the graph: mcp→host:port 62 · dbcode→store 38 ·
caddy→upstream 36 · pipeline→dest 13 · pipeline→source 13 ·
worker-route→script 12 · tunnel→origin 5 · +18 datastore relations.
247 datasets · 141 GB ingested · 88.1 GB live. Telemetry 109 / 64.4 GB · knowledge 22 / 39.2 GB · training corpora 30 / 451 MB · threat intel 22 / 35.4 MB · forensic evidence 3 / 679 MB. 24 replayable exports (11.5 GB) as hash-addressed evidence bundles.
Hybrid AI routing (§4.5): PII/hardware telemetry → local Ollama on m1; volume triage → self-hosted DeepSeek; long-context intel → Kimi K3; CN-region data → qwen3.8-max in the CN region; edge cases → gated frontier under attestation. 10–100× token-cost reduction versus premium-only routing, sensitive classes on local silicon.
Parity harness: edge vs local runtime behaviour diffed across 29 probe routes — 20/29 converged is published with a committed December 2026 remediation date (honest-ledger doctrine, §8.3).