Grafana dashboard catalog

A complete map of the DefenseClaw Grafana dashboards, which telemetry backend each board uses, how drill-downs connect, and how to interpret empty panels.

DefenseClaw provisions 18 Grafana dashboards into a DefenseClaw folder, or 19 from a source checkout (editable install), which also seeds the optional Fleet dashboard. The IoT Fleet board is the 19th: it is not in the installed package and is added with the separate Fleet assets (Fleet dashboard). Start the stack with Run Grafana locally; the links on this page assume localhost:3000.

Complete inventory

Dashboard (UID)Question it answersBackends
DefenseClaw — Overview (defenseclaw-overview)Is DefenseClaw healthy and enforcing as expected? The landing board, with links to every other board.Prometheus, Loki
DefenseClaw - Agent Activity (Live) (defenseclaw-activity)What are agents doing: prompt, model, tool, destination, session.Prometheus, Loki
DefenseClaw — Agent identity (defenseclaw-agent-identity)Which agents and instances were seen? Holds the Runtime Agent Directory that opens Agent360.Prometheus, Loki
DefenseClaw — Identity & Users (defenseclaw-identity)Who runs which agents, how verified is the identity, and which guardrail profile decided?Loki
DefenseClaw — Agent360 (defenseclaw-agent-360)What happened inside one agent, one execution, or a whole descendant tree? See Agent360.Prometheus, Loki, Tempo
DefenseClaw — Blocked Events & Analysis (defenseclaw-blocked-events)What was blocked, by which rules, and what recurs? Who was blocked on which machine, under which guardrail profile and directory group. Includes kernel denials on Linux with Tetragon.Prometheus, Loki
DefenseClaw — AI Agent Usage & Detection (defenseclaw-ai-discovery)Which AI products, skills, MCP servers, IDE plugins and dependencies exist?Prometheus, Loki, Tempo
DefenseClaw — AI Discovery Runtime Planes (defenseclaw-ai-runtime)What did those AI runtimes run on this host? Coverage first, so a blind host is not read as clean.Prometheus, Loki
DefenseClaw — Hook Connectors (defenseclaw-connectors)Which connectors are active, quiet, slow or drifting?Prometheus, Loki
DefenseClaw — Connector Detail (defenseclaw-connector-detail)Why does one connector behave differently?Prometheus, Loki
DefenseClaw — Guardrail Evaluations (defenseclaw-security)What did guardrails allow, alert, confirm or block?Prometheus, Loki
DefenseClaw — Policy decisions (defenseclaw-policy-decisions)Which policy or egress branch caused the action? Includes observe-mode would-blocks.Prometheus, Loki
DefenseClaw — HITL (Human-in-the-Loop) (defenseclaw-hitl)Where did a human approval enter, and what was the outcome?Prometheus, Loki
DefenseClaw — Findings (Rule detail) (defenseclaw-findings)Which rule fired, against what, how often?Prometheus, Loki
DefenseClaw — Scanners (Ops) (defenseclaw-scanners)Are scans healthy and producing the expected findings?Prometheus
DefenseClaw — Proxy & LLM Guard (defenseclaw-traffic)What happens on proxy and router deployments? Empty unless proxy mode is active.Prometheus, Loki, Tempo
DefenseClaw — Sandboxes (defenseclaw-sandboxes)What are OpenShell sandboxes doing, reaching and being refused? See Sandboxes board.Prometheus, Loki
DefenseClaw — Runtime & Reliability (defenseclaw-runtime)Is the process, database, exporter or destination delivery unhealthy?Prometheus, Loki

Kernel controls and your Tetragon policies

On Linux with Tetragon and DefenseClaw's sensor helper, AI Discovery Runtime Planes and Blocked Events & Analysis show kernel-layer activity. They read the plane C plane_health records, ai.runtime.kernel_policy, ai.runtime.kernel_event and enforcement.block.applied records with initiator kernel. See Tetragon on a Linux fleet.

WherePanelAnswers
Runtime PlanesFleet: kernel controls per hostTetragon version and mode per computer, whether enforce_ack approves the build, enforced and burn-in users, pauses, digests.
Runtime PlanesKernel controls (Linux)Would-block and denial counts, burn-in hours per user, DefenseClaw kernel policies loaded or removed.
Runtime PlanesYour Tetragon policies: events from AI agentsEvents and blocks by your own policies, with the agent and hook join.
Blocked EventsKernel denials, Blocked by your Tetragon policiesDenials by control, connector, account and process.

Grafana AI Discovery Runtime Planes, row Kernel controls (Linux): tiles for hosts with enforcing controls, kernel denials, would-blocks in monitor mode, agent activity with no hook decision, paused hosts and reconcile failures; charts of kernel denials by control and connector; a table of DefenseClaw kernel policies; and a table of burn-in progress per user

The Hosts with enforcing controls tile counts each computer once, using its latest plane C health record from the last 10 minutes with users_enforced above 0. The Host filter applies; the User filter does not. Five Prometheus rules cover the defenseclaw_kernel_* series:

AlertFires when
DefenseClawKernelDenialsA DefenseClaw kernel control denied an open (informational).
DefenseClawKernelAckStaleA computer's enforce_ack no longer matches its build, for 30 minutes.
DefenseClawKernelPausedLongKernel enforcement is paused for 8 hours.
DefenseClawTetragonFallbackTetragon is installed but the helper runs on cn_proc and fanotify, for 15 minutes.
DefenseClawCustomerPolicyBlocksAgentsOne of your Tetragon policies denied an AI agent (informational).

Grafana AI Discovery Runtime Planes, first row Fleet: kernel controls per host. Tiles for hosts reporting kernel controls, hosts with a stale approval, hosts on fallback, paused hosts, kernel denials and would-blocks, and a table of approval, backend, controls digest, host, mode, paused, Tetragon version, enrolled users, enforcing and in burn-in

Grafana Blocked Events and Analysis, row Kernel denials (Linux): 4 kernel denials, 1 block by an administrator's Tetragon policy, and a table of denials by connector, control, process, user and count: three reads of the SSH key by wc and one write of a shell profile by bash

Filter by user

User regex ($user) appears on Overview, Activity, Blocked Events, Policy decisions, Agent identity, Agent360, AI discovery and Identity & Users. It narrows Loki panels to records whose user.id, account name (defenseclaw.user.name) or directory principal matches the whole value, for example 1001, S-1-5-21-.*-1104, alice or alice.*. The principal is on records only when ai_discovery.include_user_principal is on. Metric panels have no per-user label and do not change. Records without a user.id drop out when you narrow it.

Who was blocked, where and under which policy

The Who was blocked, where, and under which policy row on Blocked Events lists one row per enforced guardrail decision: time, user, machine, connector, guardrail profile, the directory group whose assignment selected that profile, rule IDs and action. Profile and group are empty or none when no identity-based profile applied. Three bars count the same decisions per machine, profile and primary rule. The Machine box (host.name), Profile regex box and Rule box narrow the table, the bars and the other Loki panels on the board. Use Inspect, Data, Download CSV on the table for an audit export. The board is for administrators; it shows every user on the stack you point it at.

ID and hook surface boxes

Some boxes filter on fields only the record body carries: Root session ID regex (Activity), Agent or root ID regex, Lifecycle (drilldown) and Execution (drilldown) (Agent360), Hook surface regex (Blocked Events). Loki cannot list their values, so they take a regex. The default .+ matches everything, and table links fill them in.

Identity & Users board

Reads Loki only, with Connector and User regex boxes.

RowPanels
Who runs agentsUsers observed, Agent identities (agt-), Agent sessions (ais-), Verified principal share
Directory and sessionsUsers by directory and domain, Verified vs claimed identity, SSH vs local sessions
AgentsAgents per user, Sessions per agent
Guardrail profilesDecisions by guardrail profile and match reason, Profile selection audit, Profile digest timeline
identity.observedDirectory group counts, Recent identity.observed records

Directory shows unreported for a Linux SSSD account whose domain is not a joined realm. Principals are blank unless ai_discovery.include_user_principal is on. SSH vs local sessions counts only verified session kinds, so Windows and macOS records carry none. A new series on the digest timeline marks a reload where a profile's policy changed; compare it with defenseclaw guardrail profile explain.

Sandboxes board

The Sandboxes board (defenseclaw-sandboxes) shows OpenShell sandbox activity. It has seven boxes:

  • Environment ($environment, the gateway's deployment.environment.name, set with defenseclaw setup observability add --environment) and Host ($host, the computer's host.name). Every panel follows both, so you can scope the board to one computer and show, for example, that a Windows computer has no sandboxes.
  • Connector ($connector, the sandbox's harness connector) for the metric panels.
  • Sandbox ($sandbox, the sandbox name) for the log panels. Sandbox names are never metric labels, so the metric panels ignore it and say "(all names)" in their titles: Active (all names), Active sandboxes by connector (all names) and Egress by source and decision (metric, all names).
  • OS ($os), Driver ($driver, docker or vm) and Profile ($profile, a guardrail profile name as a regular expression, default .*) for the managed-fleet panels. Selecting a driver hides records that name none, such as host-run refusals.

The local Collector copies the sandbox name from each record's body onto the record's resource (the processors transform/sandbox-identity and groupbyattrs/sandbox-identity), so no panel parses every log line to find a sandbox. A record that names no sandbox keeps its resource unchanged.

The Environment, Host, Sandbox and OS boxes list Loki index labels, because Grafana fills a Loki box through Loki's label API. The bundled Loki indexes deployment.environment.name, host.name, os.type and defenseclaw.sandbox.name (otlp_config in loki/loki.yaml). If you import the boards into your own Loki, add the same index_label rule or the boxes list no values; panels still filter correctly. The records behind the panels, and the health codes of the integration row, are in Sandbox telemetry.

RowPanelsWhat they show
Sandbox fleetActive (all names), Sandboxes seen, Egress allowed, Egress blocked, Shadow-AI findings, Health degradedSandboxes provisioning, starting, ready or stopping now, from the defenseclaw.sandbox.active gauge (the largest value per host, so a restarted gateway is not counted twice); then counts for the range. Sandboxes seen counts a host and a name, so two computers' sandboxes with the same name count twice; pick a Host to read one computer's. Egress counts are one per decision; the end of a tunnel is not counted again. Health degraded shows two numbers: sandboxes for the sandboxes in the Sandbox box and gateway-wide for records about the whole integration.
Managed fleet: who is governed by whatGoverned sandboxes, Provisional share, Host runs refused, Sandbox unavailable, Access refused, Foreign items, Asks decided, Asks approved, Asks denied, Policy re-matches, Settings drift, OpenShell versions, Sandbox registrations without a guest, Managed sandbox events, Sandboxes by profile, Profile, match and group, Sandboxes by OS, Sandboxes by driver, OpenShell version and install modeEnterprise hosts only. Governed sandboxes counts those whose records carry an effective sandbox policy digest, so a personal OpenShell setup is not included. Provisional share is the percent of policy resolutions made on the offline floor while the directory lookup was outstanding. Settings drift counts times the daemon found OpenShell settings that differ from enterprise.openshell and refused sandboxes; zero is healthy. More than one OpenShell version means the fleet is mid-upgrade or has drifted. Per-user values are never shown.
Managed fleet: policy changes and driftPolicy re-matches by reason, Settings drift by OpenShell setting, Recent policy resolutions and re-matchesRe-matches by reason (generation for a configuration change, facts for a directory refresh, lookup_resolved for a provisional sandbox that left the offline floor), the name of each drifting OpenShell setting (never its value), and the newest resolutions with the profile and effective policy digest.
Managed fleet: refusals, asks and foreign itemsRequired-sandbox refusals by reason, Asks decided by result, Refusals by profile, Access refusals by rule, Foreign sandboxes, containers and rules, Recent refusals, unavailable sandboxes and drift, Recent asks and foreign itemsHost runs refused or unavailable by machine-readable reason (required_for, os_unsupported, no_docker_group, no_kvm, prerequisite_missing), how the managed gateway answered asks, and things found on a managed host that DefenseClaw did not create, by kind and what the managed layer did (stop, alert or remove). Containers are reported and never stopped.
LifecycleActive sandboxes by connector (all names), Phase transitions, Recent lifecycle changesThe gauge over time, the phase each change entered (creating, provisioning, starting, ready, stopping, stopped, error, deleting, deleted), and the newest changes with trigger and OpenShell condition.
Egress decisionsEgress by source and decision (logs and metric), Top blocked hosts, Top allowed hostsWho decided (openshell is OpenShell's own network boundary, dc-egress-proxy is the DefenseClaw egress proxy) and the outcome. The metric panel counts the defenseclaw.egress.events counter, which misses a series' first increment, so the log panel is the exact count.
Traffic volumeBytes up by destination, Bytes down by destination, Bytes up / down over timeSums of bytes_up and bytes_down on egress.completed records, which only the DefenseClaw egress proxy writes. OpenShell's own connections report no byte counts.
DestinationsDestinations, Recent blocked egressEach host with allowed and blocked counts and bytes moved (top 50 per column), and the newest refusals with the decision code and the process the workload claims made the request.
Findings and shadow AIFindings by kind, Shadow-AI destinations, Recent sandbox findingsFindings by kind (ocsf_finding, hook_silence, hook_tamper, large_upload, nested_repo, shadow_ai). A shadow-AI finding names an AI API that is neither the sandbox's model provider nor its harness vendor: MEDIUM when reached, LOW when refused.
Integration healthHealth degraded / restored, Recent health transitionsOpenShell health records with their stable error code. Records about the whole integration name no sandbox and show under every $sandbox selection.
Process startsProcess starts, Top started programsEmpty unless the sandbox's process tree is on (observe.process_tree, or sandbox run --process-tree). Sampling misses a process that starts and ends between two samples.

Things to know when you read it:

  • A host that did not canonicalize shows as its target reference (invalid-host, or host: plus the name for a name that starts with _).
  • Bytes stay empty for OpenShell-only traffic. Only egress.completed records from the egress proxy carry byte counts.
  • Counters miss a series' first increment. The log panels are the exact count.
  • Process names, paths and command lines are what the workload claims. The sandboxed agent chooses that text, and each destination's redaction profile decides whether it leaves the host.

Grafana Sandboxes board with the Environment, Host, Connector and Sandbox boxes, Sandbox fleet and Lifecycle rows: tiles for active sandboxes (all names), sandboxes seen, egress allowed, egress blocked, shadow-AI findings and health degraded split into sandboxes and gateway-wide, the active sandboxes by connector (all names) and phase transitions charts, and the recent lifecycle changes of the selected sandbox

Grafana Sandboxes board, Egress decisions and Traffic volume rows: egress by source and decision as logs and as a metric (all names), the top blocked host (webhook.site) and the top allowed hosts, and bytes up and down by destination and over time

Grafana Sandboxes board, Process starts row: process starts over time and the top started programs, filled because the sandbox ran with the process tree on

How the drill-down works

pick a connector
pick an agent
pick a trace
Overviewhealth and KPIs
Hook Connectors
Agent identity
Blocked Events
AI Agent Usage& Detection
Connector Detail
Agent360
Policy decisions
Runtime Planeswhat actually ran
Tempo waterfallone selected trace

List view for small screens. Use the expand button to open the drawing.

  1. Overviewhealth and KPIs
    • Hook Connectors
    • Agent identity
    • Blocked Events
    • AI Agent Usage
  2. AI Agent Usage& Detection
    • Runtime Planes
  3. Blocked Events
    • Agent360
    • Policy decisions
  4. Agent identity
    • pick an agentAgent360
  5. Hook Connectors
    • pick a connectorConnector Detail
  6. Runtime Planeswhat actually ran
  7. Policy decisions
  8. Agent360
    • pick a traceTempo waterfall
  9. Connector Detail
  10. Tempo waterfallone selected trace
The main investigation paths, each one a link on the board it starts from. Every board also links back to Overview, and Agent360 opens a selected trace in its Tempo waterfall.

Agent and trace IDs are links. An Agent ID opens Agent360 with the time range kept. A Trace ID sets the trace variable on the same board and fills the Tempo waterfall without losing the connector, agent, lifecycle or execution filters.

Which backend owns each answer

BackendBest forEmpty means
PrometheusTotals, rates, quantiles, current state, inventoryThe series does not exist in the range. Counter totals use the selected range, so a finished run stays visible.
LokiLifecycle graph, chronology, inputs and outputs, reasons, decisionsNo matching event in the range and filters.
TempoOne request and its span timingNo valid Trace ID selected, or no matching spans.

Dashboards add no redaction, masking or field hiding. They show every field that reaches Loki, Tempo or Prometheus, including content when the producer supplied it. Set a redaction profile on the local-observability destination to reduce what they receive.

Read empty values correctly

DefenseClaw does not invent telemetry.

DisplayMeaning
0The signal is instrumented and the range has zero matching events.
No dataThe panel is conditional (approvals, findings, proxy-only traffic) and nothing matches the current filters.
Not reportedThe event exists, but the connector or provider did not send that optional field, most often tokens or cost. It is not zero and is never estimated.

If panels stay empty while Loki and Tempo have events, check the destination name. The boards read the short local-observability-v1 metric labels, which the gateway exports only to the OTLP destination named local-observability. Any other name gets canonical labels, and connector- and scanner-filtered panels show nothing. If a global panel is empty while raw events are visible, widen the time range first.

If rare events (a single block or finding) read 0 in Prometheus panels, counters are exported as deltas and rate() needs two samples. The Findings and Blocked Events tiles count Loki log records and are exact. For the others, set observability.metric_policy.temporality: cumulative (Observability reference).

Source of truth and upgrades

The optional DefenseClaw IoT Fleet (defenseclaw-fleet) dashboard ships in the separate defenseclaw-fleet-assets.tar.gz release asset, alongside the authenticated Fleet scrape templates. See Install the Edge Connector for Fleet setup. The core wheel excludes these optional Fleet assets. A source checkout (editable install) seeds the Fleet dashboard from bundles/ along with the other 18.

The dashboard sources live in bundles/local_observability_stack/grafana/dashboards/. up refreshes the installed dashboards on every start; pass --no-refresh-config to keep your edits. Edits made in the Grafana UI are not written back to disk, so copy a changed board out with Dashboard → JSON Model.