AGENT STACK MAP
backend: -- network >_ terminal
STACK: evaline.network/stack /network /terminal | github @evaline-online
OFFLINE-CAPABLE - NO EXTERNAL CDN - STATIC HTML
+-- evaline.network/stack ----------------+ | LLM AGENT STACK // TOPOLOGY . MATRIX | | LIVE TELEMETRY | +----------------------------------------+

EvaLine / EvaBot LLM Agent Stack

Single-page map of the compute + micro cluster, the OmniRoute model gateway, every agent runtime, the shared MCP bus, persistent memory stores and outbound external services. Rendered as hand-drawn SVG with no libraries; live panels poll same-origin backend endpoints every 30 seconds.

[01] STACK SUMMARY
static counts (verified snapshot)
OmniRoute gateway
36
models served at :20128
ModelRegistry
94
62 free / 32 paid
MCP core servers
8
21 in the full suite
Agent runtimes
7
agy . 2.0 . IDE . opencode . kilo . hermes . openclaw
Cluster VMs
2
compute (eu) + micro (us)
[02] INTERACTIVE TOPOLOGY
hover to trace - click a node for details
HTTPS/QUIC WireGuard proxy :3000 LLM /v1 OpenAI-compatible inference (agy is the exception) MCP stdio/SSE (shared bus) graph + sqlite kb / chroma / chat provider API gh / n8n / gemini OAuth / Cloud Code RAG (NotebookLM) OUTBOUND EGRESS BUS BROWSER evaline.network Browser / Client EVALINE-MICRO-VM Caddy HTTP/3 ingress 136.114.26.252 evaline-micro-vm NGINX :80 compute nginx :80 EVABOT-BACKEND :3000 Node.js ModelRegistry 94 evabot-backend :3000 OMNIROUTE / LiteLLM 127.0.0.1:20128/v1 single gateway - 36 models OmniRoute :20128 HERMES :8642 API server / hermes-agent Hermes :8642 OPENCLAW :18789 gateway OpenClaw :18789 AGY (ANTIGRAVITY) Google Cloud Code OAuth agy OPENCODE via OmniRoute OpenCode KILOCODE via OmniRoute KiloCode QWEN via OmniRoute Qwen MCP BUS - 8 CORE filesystem git github fetch memory sequential-thinking sqlite context7 21 servers in the full suite MCP bus MEMORY / DATA ~/.mcp/memory.json [EMPTY] ~/.mcp/sqlite.db ChromaDB (1075 vectors) knowledge.db / chat-history.db Memory / Data EXTERNAL SERVICES // OUTBOUND EGRESS GEMINI API 403 generateContent Gemini API GITHUB gh API + MCP server GitHub GCLOUD / VERTEX billing account closed gcloud / Vertex AI N8N :5678 workflows / webhooks n8n :5678 NOTEBOOKLM not authenticated NotebookLM
healthy / verified known issue (see matrix) highlighted connection directed edge live data path
[select a node]
Click any node in the topology to inspect its role, status and outgoing connections. Hovering traces a node's edges.
[03] CONNECTION MATRIX
runtime x layer
Runtime Edge ingress LLM gateway :20128 Cloud Code OAuth MCP bus (8) Memory / data External APIs
Hermes:8642 VIA MICROCaddy -> nginx -> backend OKhermes-agent route N/A OK8 servers WARNmemory.json empty OKgithub / fetch
OpenClaw:18789 VIA MICRO OKvia :20128 N/A OKmcp.servers WARNmemory.json empty OK
agy (Antigravity)CLI / 2.0 / IDE N/Alocal CLI BYPASSindependent of OmniRoute OKOAuth OKflat mcpServers WARN WARNVertex / billing closed
OpenCodesession runtime N/A OK N/A OK WARN OK
KiloCodekilo.jsonc N/A OK N/A OK WARN OK
Qwenruntime N/A OK N/A OK WARN WARNpolicy not mirrored
evabot-backend:3000 OKnginx :80 OK4 providers N/A SHARED WARNgraph empty WARNgemini 403
OmniRoute:20128 N/A SELF36 models N/A N/A N/A WARNgemini 403 / CF-Groq quota
Gemini generateContent 403BLOCKING
The Gemini API key is valid and lists 50 models (HTTP 200), but the GCP project is denied for generateContent (403). Every omni/gemini-* route therefore runs on fallback providers. Needs a new GCP project / key.
Cloudflare / Groq free quotasWARN
Free-tier providers behind OmniRoute (for example omni/cf-llama-3.3-70b and Groq-hosted models) are rate/quota limited and can throttle or fail under load. Fallback chains absorb most of this but add latency.
NotebookLM not authenticatedWARN
Grounded RAG is offline. As of 2026-09-15 the session is unauthenticated: run notebooklm_setup_auth, then add_notebook and select_notebook. Intended default notebook: Antigravity.
memory.json empty / unusedWARN
The persistent knowledge graph at ~/.mcp/memory.json exists but is empty, so graph memory is inactive. SQLite (~/.mcp/sqlite.db), ChromaDB and the backend databases remain populated.
Vertex AI / Agent Engine / GCS / Secret ManagerBLOCKED
These GCP services are blocked by a closed (delinquent) billing account. Restoring billing is required before they can be used.
Verified healthyOK
Backend :3000, nginx :80, OmniRoute :20128, Hermes :8642 (auth-gated), OpenClaw :18789, n8n :5678 and the WireGuard mesh between the two VMs responded as expected in the canonical stack check.
[04] LIVE BACKEND PANELS
same-origin fetch - refresh 30s
last updated: -- | next poll in 30s
GET /api/health
awaiting first poll - shows status, version, uptime, memory, CPU/load, node telemetry and database counts.
GET /api/models
awaiting first poll - shows ModelRegistry totals (free / paid), categories, default model and fallback chain length.
GET /api/logs/recent
awaiting first poll - shows the most recent backend request log lines with level and status.
GET /api/security/status
awaiting first poll - shows blocked IPs, suspicious patterns and total tracked addresses.
GET /api/stack/status
awaiting first poll - probes backend, voice, face, n8n, openclaw, hermes, omniroute, nginx and the micro mesh hop with per-target status, latency and ok flag.