Configuration

All configuration is via environment variables. Everything has a default — both binaries start with zero config against a local kubeconfig.

Settings UI

Most agent parameters can also be changed at runtime on the Settings page (sidebar → System → Settings): LLM provider, model, Ollama host, Anthropic API key (write-only — never displayed), review interval, and the periodic review on/off switch.

How it works:

• Changes apply immediately in the running agent — no pod restart, the insight history is kept, and a fresh review runs right away.
• The agent also persists the change to a settings table in its own SQLite file (the same one insight history uses), encrypted — not plaintext or base64. This works identically in Docker and k8s mode, as long as INSIGHT_DB_PATH points at a persistent volume (the default in both docker-compose.yml and the Helm chart / raw manifests). No volume, no key file, no persistence — the UI tells you which case you're in (persistent: true/false) after every save.
• Once anything has ever been saved, it's permanent: the deployment's env vars / Secret only matter on the agent's very first boot (a genuinely empty database). A later helm upgrade --set, kubectl edit, or manifest re-apply has no effect — the Settings UI is the only way to change something after that first boot. See security.md.

Server parameters (port, agent URL, RBAC) are deployment-level and remain read-only in the UI — change them in the manifests / compose file.

Server (cmd/server)

Variable	Default	Description
PORT	8080	HTTP listen port
AGENT_URL	http://localhost:8090	Base URL of the agent service; /api/v1/agent/* is proxied here
KUBECONFIG_PATH	(auto)	Explicit kubeconfig path. When unset: in-cluster config if running in a pod, else standard kubeconfig rules (KUBECONFIG env, ~/.kube/config)
STATIC_DIR	(auto)	Directory of the built SPA. When unset: web/dist if present, else API-only (dev mode)
LOG_LEVEL	info	debug, info, warn, error
LOG_FORMAT	text	text or json

Agent (cmd/agent)

Variable	Default	Description
PORT	8090	HTTP listen port
LLM_PROVIDER	ollama	ollama (local, free) or a cloud provider: anthropic, openai, deepseek, gemini
LLM_MODEL	(per provider)	Model ID. Defaults: qwen3:0.6b (ollama), claude-opus-4-8 (anthropic), gpt-5.1 (openai), deepseek-chat (deepseek), gemini-2.5-flash (gemini). The k8s manifests set qwen3:0.6b to fit small clusters
ANTHROPIC_API_KEY	—	Required when LLM_PROVIDER=anthropic
OPENAI_API_KEY	—	Required when LLM_PROVIDER=openai
DEEPSEEK_API_KEY	—	Required when LLM_PROVIDER=deepseek
GEMINI_API_KEY	—	Required when LLM_PROVIDER=gemini
OLLAMA_HOST	http://localhost:11434	Ollama base URL (ollama provider only)
AGENT_REVIEW_INTERVAL	5m	How often the monitor reviews the cluster (Go duration, min 30s)
AGENT_MONITOR_ENABLED	true	Set false to disable the periodic review loop (queries still work)
NOTIFICATIONS_ENABLED	false	Send Discord alerts when the cluster status changes (healthy→warning→critical and recoveries). Never per-review
DISCORD_WEBHOOK_URL	—	Discord webhook to post alerts to. Write-only in the API/UI
SLACK_WEBHOOK_URL	—	Slack incoming webhook (https://hooks.slack.com/...). Write-only in the API/UI
TEAMS_WEBHOOK_URL	—	Microsoft Teams Workflows webhook (Adaptive Card payload). Write-only in the API/UI
NOTIFY_MIN_SEVERITY	warning	warning = alert on any degradation; critical = only page for outages (recoveries from qualifying states still notify)
REPORT_ENABLED	false	Send a daily digest of the last 24h (reviews, incidents, remediation proposals, LLM usage) to the same webhooks. Pure reporting from stored data — no extra LLM calls
REPORT_TIME	08:00	When the daily report goes out — HH:MM, 24-hour, UTC
INSIGHT_DB_PATH	(empty)	SQLite file for persistent review history (e.g. /data/insights.db). Empty = in-memory only, history lost on restart. The k8s manifests set this and mount a PVC
INSIGHT_RETENTION_DAYS	90	Reviews older than this are pruned from the database. Seeds the first-boot default; changeable at runtime from the Settings page (1–3650 days)
PROMETHEUS_URL	(empty)	Prometheus base URL for the agent's metrics tools (usage, throttling). The k8s manifests point this at the bundled Prometheus; set your own to reuse an existing one. Empty = metrics tools disabled
KUBECONFIG_PATH	(auto)	Same resolution as the server
LOG_LEVEL / LOG_FORMAT	info / text	Same as the server

LLM provider setup

Ollama (local models — the default)

1. Get an Ollama endpoint:
   • k8s mode: deploy/k8s/05-ollama.yaml runs it in-cluster and auto-pulls qwen3:0.6b on first boot — nothing to do.
   • Docker mode: docker compose --profile ollama up, then docker compose exec ollama ollama pull qwen3:0.6b once.
   • Local dev: install Ollama (https://ollama.com) and ollama pull qwen3:0.6b.
2. The model must be tool-calling-capable — the query engine requires function calling. Known-good: qwen3 / qwen3:0.6b, llama3.1, mistral-nemo.
3. Override with LLM_MODEL and OLLAMA_HOST as needed.

Notes:

• Small local models (like the 0.6b default in the k8s manifests) are noticeably weaker at the structured review and multi-step tool use. If reviews come back as status: "error" with a parse failure or queries wander, move up to qwen3 (8B) or switch to anthropic.
• First request after a pull loads the model into memory and can take a minute; the agent tolerates up to 5 minutes per LLM call.

Cloud providers (Anthropic, OpenAI, DeepSeek, Gemini)

All four work the same way — pick the provider, provide its API key (easiest: Settings page, which stores it write-only and persists it encrypted to the agent's own database), optionally pick a model from the dropdown:

Provider	Get a key at	Default model	Notes
anthropic	platform.claude.com	claude-opus-4-8	Strongest agentic tool use in our testing
openai	platform.openai.com	gpt-5.1
deepseek	platform.deepseek.com	deepseek-chat	Very low cost; deepseek-reasoner for harder analysis
gemini	aistudio.google.com	gemini-2.5-flash	Served via Google's OpenAI-compatible endpoint

The model dropdowns show a curated list; any valid model ID for that provider also works (the lists are a convenience, not an allowlist). The agent requires models that support function calling — all listed defaults do.

Cost control (anthropic)

The monitor loop calls the LLM once per interval, 24/7. With the default 5m interval that is ~8,600 calls/month; each call sends a compact snapshot (typically 1–3K tokens) and receives a short JSON verdict. Levers, in order of impact:

1. AGENT_REVIEW_INTERVAL=15m (or more) — proportional cost reduction
2. AGENT_MONITOR_ENABLED=false — on-demand queries only
3. LLM_MODEL — a smaller/cheaper Claude model for the loop