# Production Checklist

Review each section before exposing chmonitor to a team or the internet.

## ClickHouse user and grants

- [ ] Create a dedicated monitoring user — do not use `default` or an admin account.
- [ ] Grant `SELECT` on `system.*` tables only. The dashboard is read-only by default.
- [ ] If you enable the kill-query action, also grant `KILL QUERY`.
- [ ] Set `CLICKHOUSE_MAX_EXECUTION_TIME` to a safe limit (e.g. `30`) so runaway dashboard queries stop quickly.
- [ ] Store `CLICKHOUSE_PASSWORD` in a secret (Docker secret, k8s Secret, Cloudflare Worker secret, Vercel env) — never in source code or a ConfigMap.
- [ ] Prefer HTTPS for ClickHouse Cloud and any public endpoint. Keep private endpoints behind your own network.

## Authentication — do not leave fully public

- [ ] Choose an auth provider: `none` is only safe on a private network with no agent or MCP access.
- [ ] If you use `CHM_AUTH_PROVIDER=none`, restrict network access so the app is not reachable from the internet.
- [ ] For team deployments, use Clerk or a trusted-header proxy provider.
- [ ] Set `CHM_API_KEY_SECRET` whenever you expose `/api/mcp`. MCP clients send long-lived tokens — require an API key. (The MCP endpoint returns 401 by default; set `CHM_MCP_PUBLIC=true` only on isolated private networks.)
- [ ] Rotate `CHM_API_KEY_SECRET` and `CHM_PROXY_AUTH_SECRET` if they are ever exposed. Redeploy after rotating.
- [ ] Never put `CLERK_SECRET_KEY`, `LLM_API_KEY`, or `CHM_API_KEY_SECRET` in a `VITE_*` or `NEXT_PUBLIC_*` variable — those are baked into browser JS.

## Feature permissions — gate sensitive features

- [ ] Gate the agent behind authentication: `CHM_FEATURE_AGENT_ACCESS=authenticated`.
- [ ] Gate settings: `CHM_FEATURE_SETTINGS_ACCESS=authenticated` or `CHM_FEATURE_SETTINGS_ENABLED=false`.
- [ ] Gate MCP: `CHM_FEATURE_MCP_ACCESS=authenticated`.
- [ ] Disable features your team does not use: `CHM_DISABLED_FEATURES=actions,peerdb`.
- [ ] Review which features are visible to unauthenticated users. Default is all public.

## AI agent

- [ ] Keep `AGENT_ENABLE_CONTROL_TOOLS=false` (default) unless you trust all authenticated users to kill queries and optimize tables.
- [ ] Keep `LLM_API_KEY` server-side only — never in a client env var.
- [ ] Set `AGENT_API_TOKEN` if you want to gate programmatic access to `/api/v1/agent`.
- [ ] The agent uses the configured ClickHouse user's grants. A read-only user limits what the agent can do.
- [ ] Set spending / rate limits on your LLM provider to cap runaway agent costs.

## Health alerting

- [ ] Set `CRON_SECRET` to protect the `/api/cron/health-sweep` endpoint.
- [ ] Set `HEALTH_ALERT_ENABLED=true` and `HEALTH_ALERT_WEBHOOK_URL` (Slack or Discord) for proactive alerts.
- [ ] Choose `HEALTH_ALERT_MIN_SEVERITY=warning` or `critical` depending on alert volume.
- [ ] Verify the cron trigger reaches the endpoint (Cloudflare Cron Trigger, k8s CronJob, or external cron).

## Query timeout and pool

- [ ] Set `CLICKHOUSE_MAX_EXECUTION_TIME` below your infrastructure's request timeout (30 s is a safe default).
- [ ] Tune `CLICKHOUSE_POOL_SIZE` to match expected concurrent dashboard users.

## Conversation store backups

- [ ] If using a server-side conversation store (postgres, clickhouse, D1), include it in your backup schedule.
- [ ] For D1 on Cloudflare, use `wrangler d1 export` to back up.
- [ ] For postgres, use your database provider's backup feature.
- [ ] `memory` and `local` (browser localStorage) stores are ephemeral — no backup needed, but history is lost on restart or when browser data is cleared.

## Network and deployment

- [ ] Verify the app can reach ClickHouse from the deployment runtime, not just from your laptop.
- [ ] Use HTTPS in front of chmonitor when exposing to a team (TLS termination at nginx, Cloudflare, or your load balancer).
- [ ] Record where secrets are stored and who can rotate them.
- [ ] Keep a rollback path: previous Docker tag, Helm release revision, or Worker version.
- [ ] Watch app logs after deploy for ClickHouse connection errors or auth failures.