██████╗ ██╗ ██╗ ███╗ ███╗██╗ ██╗ ██████╗ ██████╗ █████╗ ██████╗ ██╗ ██╗ ██╗ ██╔═══██╗██║ ██║ ████╗ ████║╚██╗ ██╔╝ ██╔════╝ ██╔══██╗██╔══██╗██╔══██╗██║ ██║ ██║ ██║ ██║███████║ ██╔████╔██║ ╚████╔╝ ██║ ███╗██████╔╝███████║██████╔╝███████║ ██║ ██║ ██║██╔══██║ ██║╚██╔╝██║ ╚██╔╝ ██║ ██║██╔══██╗██╔══██║██╔═══╝ ██╔══██║ ╚═╝ ╚██████╔╝██║ ██║ ██║ ╚═╝ ██║ ██║ ╚██████╔╝██║ ██║██║ ██║██║ ██║ ██║ ██║ ╚═════╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═╝ ╚═╝ Because dashboards suck, but graphs do not!
Install dependencies:
bun installRun the frontend and backend dev servers:
bun run dev
The chat pane uses the official Codex package and SDK from the Bun server.
- Start the app with
bun run dev. - Click
Connect Codexin the chat pane. - Open the displayed verification URL and enter the one-time code.
- When Codex reports authenticated, send messages from the chat input.
Device auth requires device-code login to be enabled for your ChatGPT account or workspace. Codex stores the resulting session in its normal local auth cache, such as ~/.codex/auth.json or your OS credential store.
The Bun server can query Grafana Cloud metrics through the Prometheus HTTP API. Configure these environment variables before starting the server:
export GRAFANA_PROMETHEUS_URL="https://<your-prometheus-endpoint>/api/prom"
export GRAFANA_PROMETHEUS_USER="<grafana-cloud-instance-id>"
export GRAFANA_PROMETHEUS_TOKEN="<grafana-cloud-access-policy-token>"Use the Prometheus endpoint from your Grafana Cloud stack details, not the Grafana UI URL. It usually looks like https://prometheus-prod-<region>.grafana.net/api/prom.
For Grafana Cloud Prometheus auth, use basic auth credentials from the metrics data source details:
GRAFANA_PROMETHEUS_USER is the metrics instance ID.
GRAFANA_PROMETHEUS_TOKEN is a Grafana Cloud access policy token with metrics:read. Grafana UI service-account tokens do not work for the hosted Prometheus endpoint.
GRAFANA_PROMETHEUS_URL can also point directly at an /api/v1 Prometheus-compatible base URL, or a copied /api/v1/query URL. If no user is set, the server sends the token as a bearer token; Grafana Cloud commonly expects basic auth with the instance ID as the user and the token as the password.
Direct endpoints:
curl -X POST http://localhost:3000/api/grafana/prometheus/query \
-H 'content-type: application/json' \
-d '{"query":"up"}'
curl -X POST http://localhost:3000/api/grafana/prometheus/query_range \
-H 'content-type: application/json' \
-d '{"query":"up","start":"2026-05-13T11:00:00Z","end":"2026-05-13T12:00:00Z","step":"60s"}'
curl -X POST http://localhost:3000/api/grafana/prometheus/series \
-H 'content-type: application/json' \
-d '{"match":"up","start":"2026-05-13T11:00:00Z","end":"2026-05-13T12:00:00Z"}'
curl -X POST http://localhost:3000/api/grafana/prometheus/label_values \
-H 'content-type: application/json' \
-d '{"label":"__name__","start":"2026-05-13T11:00:00Z","end":"2026-05-13T12:00:00Z"}'
curl -X POST http://localhost:3000/api/grafana/prometheus/label_values \
-H 'content-type: application/json' \
-d '{"label":"__name__","match":"{__name__=~\".*(cpu|memory|mem|disk|filesystem|node|host).*\"}","start":"2026-05-13T11:00:00Z","end":"2026-05-13T12:00:00Z"}'If an instant query such as up returns no series but you expect data, try a lookback query like max_over_time(up[5m]) or use /series with match: "up" and a recent start/end window. Instant Prometheus queries only return currently active series at the evaluation timestamp.
The Codex chat endpoint can also ask Bun to run these Prometheus queries internally when a user asks for metrics or time-series data.
OMG keeps recent chat turns plus a compressed conversation summary as Codex context. When estimated memory reaches 70% of the configured context budget, the server folds older turns into the summary, retains the most recent turns verbatim, resets the Codex thread, and sends the summary with future prompts. Tune this with CODEX_CHAT_CONTEXT_LIMIT_TOKENS, CODEX_CHAT_COMPRESSION_THRESHOLD_RATIO, and CODEX_CHAT_RETAINED_HISTORY_ENTRIES.
OMG also uses an evidence-first answer contract for observability responses. Codex must return structured observations, interpretation, unknowns, and next checks. The server validates that Prometheus-backed observations and canvas panels cite PromQL queries actually executed in the current turn, rejects overconfident root-cause language, and asks Codex to repair unsupported answers before returning them. Empty Prometheus results are treated as missing/no matching series, not as observed zero.
Codex can also instruct the frontend canvas to render metrics. POST /api/chat may return a canvas array with items like:
[
{ "type": "single_stat", "label": "Healthy", "dataType": "bool", "value": true, "query": "max_over_time(up[5m])" },
{ "type": "single_stat", "label": "CPU", "dataType": "percentage", "value": 82.4, "query": "100 - avg(rate(node_cpu_seconds_total{mode=\"idle\"}[5m])) * 100" },
{
"type": "graph",
"title": "Requests",
"categories": ["10:00", "10:01"],
"query": "sum(rate(http_requests_total[1m]))",
"series": [{ "name": "requests", "data": [12, 18] }]
},
{
"type": "stacked_timeseries",
"title": "Requests by Status",
"categories": ["10:00", "10:01"],
"series": [
{ "name": "2xx", "data": [120, 132] },
{ "name": "5xx", "data": [2, 5] }
]
},
{ "type": "bar", "title": "CPU by Host", "categories": ["host-a", "host-b"], "series": [{ "name": "cpu %", "data": [64, 82] }] },
{ "type": "gauge", "title": "Disk Used", "value": 87, "min": 0, "max": 100, "unit": "%", "thresholds": { "warning": 75, "critical": 90 } },
{ "type": "table", "title": "Top Filesystems", "columns": ["host", "mount", "used %"], "rows": [["host-a", "/", 87], ["host-b", "/var", 72]] },
{
"type": "heatmap",
"title": "Host Alerts",
"columns": ["Disk", "Memory", "CPU"],
"query": "ALERTS{alertstate=\"firing\"}",
"rows": [
{ "label": "host-a", "values": ["ok", "warning", "ok"] },
{ "label": "host-b", "values": ["critical", "ok", "unknown"] }
]
},
{
"type": "alert_list",
"title": "Active Alerts",
"alerts": [{ "name": "DiskAlmostFull", "severity": "critical", "state": "firing", "target": "host-a", "summary": "Root filesystem is above 90%" }]
},
{
"type": "histogram",
"title": "Request Duration Buckets",
"buckets": [{ "label": "0.1s", "value": 42 }, { "label": "0.5s", "value": 18 }]
},
{
"type": "event_timeline",
"title": "Recent Events",
"events": [{ "time": "2026-05-13T12:00:00Z", "title": "Deploy started", "severity": "ok", "description": "version abc123" }]
},
{
"type": "markdown",
"title": "Notes",
"content": "Check host-a first; it has the highest disk pressure."
}
]Panels with a query value render a hoverable info icon showing the PromQL that produced the panel.
For multi-host questions, Codex is instructed to combine aggregate and per-host views: aggregate single stats, per-host CPU/memory/disk graphs, host alert heatmaps when alert data is available, and a prose conclusion that calls out the most important risks.
Codex can chain multiple Grafana tool calls before answering. This lets it discover metric names and labels first, then run candidate PromQL queries instead of guessing node-exporter metric names.
Graph panels should be based on Prometheus range queries. When the user asks for utilization, trends, host health panels, diagrams, or time-series views and no time window is specified, Codex defaults to the last 1 hour with step: "60s". Instant queries are reserved for single-stat snapshots such as current values, counts, and booleans.
Prometheus matrix results from query_range are mapped into canvas graphs by using sample timestamps as categories, series labels such as host/instance/device as series[].name, and numeric sample values as series[].data.
Build the frontend and backend:
bun run buildStart the Bun server:
bun run startThe production server serves dist/client and exposes GET /api/health, Codex auth routes under /api/codex/auth/*, Grafana Prometheus routes under /api/grafana/prometheus/*, and POST /api/chat.