Skip to main content
Glama

OPNsense / pfSense Firewall Operations

Audit availability: Real writes require a committed pending audit row and stop before contacting the platform if storage is unavailable. Reads and previews may continue with a warning; new audit rows store outcome summaries rather than full responses.

Governed, audited AI-ops for OPNsense and pfSense firewalls — for AI agents (via MCP) and humans (via CLI).

Disclaimer: Community-maintained open-source project. Not affiliated with, endorsed by, or sponsored by the OPNsense project, Deciso, Netgate, or the pfSense project. OPNsense, pfSense and Netgate are trademarks of their respective owners. MIT licensed.

firewall-aiops speaks to two firewall platforms behind one MCP server — OPNsense (REST API under /api/..., API key+secret via HTTP Basic auth) and pfSense (REST API v2 under /api/v2/... from the pfSense-pkg-RESTAPI package, API key via an X-API-Key header) — with the same tools working on both. Each target in the config names its own platform; a name-keyed platform registry selects the API shape (auth + resource paths), so an agent never has to know which firewall it is talking to.

Every tool runs through a built-in governance harness (vendored, zero external dependency): audit log, token/call budget with runaway circuit-breaker, descriptive risk-tier labelling, undo-token recording, and prompt-injection sanitisation.

Why this exists

  • One server, both firewalls — OPNsense and pfSense in a mixed estate, spoken to through identical tool names. Adding a third firewall later is a new platform descriptor, not a rewrite.

  • Read the whole firewall — firmware/health, interfaces & gateways, filter rules (with hit counts and state table), NAT (port-forward / outbound / 1:1), aliases, VPN (WireGuard / OpenVPN / IPsec), DHCP leases & reservations, and the firewall log.

  • Flagship RCA analyses — transparent heuristics that show their numbers, never a black-box verdict: gateway_health_rca (WAN loss/latency/down → cause + action), rule_hit_and_shadow_analysis (never-hit + shadowed/redundant rules), and blocked_traffic_rca (top blocked sources/ports → scan / brute-force / probe).

  • Governed writes — toggle a rule, add/remove an alias entry (reversible, undo-recorded from the fetched before-state), flush states, restart a service, and the "make it live" commit (apply_changes / reconfigure) and reboot at risk=high with a dry-run preview. Every write, reversible or not, lands an audit row.

Related MCP server: observability-aiops

What this tool does, and does not, decide

It delivers firewall operations — reads and writes — accurately and efficiently, and records every one of them. It does not decide whether a write is allowed to happen. That is the agent's judgement, or the permission of the account you connect it with: give the OPNsense/pfSense API user a read-only role, and the writes fail at the server — the place that actually owns the permission.

So there is no read-only switch, no policy file, no approval gate to configure. The one thing the tool guarantees is that nothing is silent: every call, over MCP and over the CLI alike, lands an audit row in ~/.firewall-aiops/audit.db, and reversible writes still capture their before-state and record an inverse where one exists.

Each tool declares a risk_level, matched to its [READ]/[WRITE] documentation tag, and carried into the audit row as a descriptive tier — so a reviewer can see at a glance that a row was a high-risk write. It is a label, not a gate.

Running a smaller / local model? See agent-guardrails.md — it lists the guardrails this tool now enforces for you (so you don't spend prompt budget restating them) and gives a ready-made system prompt for what's left.

Tool inventory (35 tools)

Domain

Tools

#

Kind

System

firmware_status, health_status, interface_status, gateway_status

4

read

Rules

list_rules, rule_detail, rule_stats, rule_states

4

read

NAT

nat_port_forwards, nat_outbound, nat_one_to_one

3

read

Aliases

list_aliases, alias_entries

2

read

VPN

wireguard_status, openvpn_sessions, ipsec_sas

3

read

DHCP

dhcp_leases, dhcp_static_mappings

2

read

Diagnostics

firewall_log, states_table, top_talkers

3

read

Flagship analyses

gateway_health_rca, rule_hit_and_shadow_analysis, blocked_traffic_rca

3

read

Writes

toggle_rule, add_alias_entry, remove_alias_entry, kill_states, restart_service

5

write (med)

Writes

apply_changes, reconfigure, reboot

3

write (high)

Undo

undo_list, undo_apply

2

read / write

Reversible writes record an inverse undo descriptor built from the real fetched before-state (toggle_rule restores the rule's prior enabled flag; alias add/remove invert). apply_changes / reconfigure / reboot are high-risk with dry_run; reboot is irreversible (audit only).

Install

uv tool install firewall-aiops        # or: pipx install firewall-aiops

Quick start

As a Claude Code plugin

One install gives an agent both the skill and the MCP server:

/plugin marketplace add AIops-tools/marketplace
/plugin install firewall-aiops@aiops-tools

The MCP server is fetched with uv and pinned to the package version this plugin declares, so an audit row can be traced back to the code that wrote it. Credentials are still configured with firewall-aiops init — see below.

As an OpenClaw plugin

The same bundle is published on ClawHub, where one install delivers the skill and its MCP server together:

openclaw plugins install clawhub:@zw008/firewall-aiops
openclaw skills info firewall-aiops          # expect: Visible to model: yes

Restart the OpenClaw gateway afterwards so it loads the plugin. The MCP server is fetched with uv, pinned to this exact release, so uvx has to be on PATH — without it the skill still installs but reports Visible to model: no. Credentials are configured exactly as below.

As a CLI or standalone MCP server

firewall-aiops init                     # wizard: pick platform (opnsense/pfsense) + store the secret (encrypted)
firewall-aiops doctor                   # verify config, secrets, and connectivity
firewall-aiops overview                 # one-shot: version + gateway/interface health + rule count
firewall-aiops rules list               # list filter rules
firewall-aiops rules toggle <uuid> --disable   # dry-run + double-confirm governed write
firewall-aiops log --action block -n 50 # recent blocked traffic

Run the MCP server (stdio) for an agent:

firewall-aiops mcp                      # or: firewall-aiops-mcp

MCP client config

{
  "mcpServers": {
    "firewall-aiops": {
      "command": "uvx",
      "args": ["--from", "firewall-aiops", "firewall-aiops-mcp"],
      "env": { "FIREWALL_AIOPS_MASTER_PASSWORD": "your-master-password" }
    }
  }
}

Configuration

~/.firewall-aiops/config.yaml (non-secret connection details only):

targets:
  - name: fw1
    platform: opnsense       # opnsense | pfsense
    host: 192.0.2.1
    port: 443
    username: <opnsense-api-key>   # OPNsense API key (unused for pfSense)
    verify_ssl: true         # the default; false ONLY for self-signed lab certs
  - name: edge
    platform: pfsense
    host: 192.0.2.2
    verify_ssl: true
    # scheme: http           # https is the default; http ONLY for a GUI behind a
    #                        # proxy that terminates TLS for it

The secret — the OPNsense API secret (paired with the key for HTTP Basic auth) or the pfSense API key — is stored encrypted in ~/.firewall-aiops/secrets.enc (Fernet + scrypt-derived key), never plaintext on disk. Set it with firewall-aiops secret set <target> or the init wizard. The store is unlocked by a master password from FIREWALL_AIOPS_MASTER_PASSWORD (non-interactive/MCP/CI) or an interactive prompt (CLI on a TTY). A legacy plaintext env var FIREWALL_<TARGET>_SECRET is honoured as a fallback (migrate with firewall-aiops secret migrate).

Governance

Every MCP tool is wrapped by @governed_tool:

  • Audit — every call is logged to ~/.firewall-aiops/audit.db (tool, params with secrets redacted, status, duration, risk tier, approver, rationale).

  • Budget / runaway guard — per-process token/call caps and a repeat-call circuit breaker (FIREWALL_MAX_TOOL_CALLS, FIREWALL_RUNAWAY_MAX, …).

  • Risk-tier labelling — each tool's declared risk_level is recorded on its audit row as a descriptive tier (a label, not a gate); there is no read-only switch, policy file, or approval gate. FIREWALL_AUDIT_APPROVED_BY and FIREWALL_AUDIT_RATIONALE are optional audit annotations, recorded when set but never required.

  • Undo recording — reversible writes record an inverse descriptor to ~/.firewall-aiops/undo.db from the fetched before-state (recording only; an external orchestrator executes it).

  • Sanitisation — all firewall-returned text is bounded + injection-sanitised before it reaches the agent.

Platform support & verification status

  • Platforms: OPNsense (REST API) and pfSense (REST API v2, pfSense-pkg-RESTAPI).

  • Test coverage: behaviour is validated against mocked OPNsense/pfSense JSON responses — every module imports, every MCP tool carries the governance marker, the RCA heuristics are unit-tested against synthetic telemetry, and reversible writes are asserted to record the correct inverse undo descriptor. The concrete REST paths are modelled from each project's public API and have not yet been exercised against a live firewall. See docs/VERIFICATION.md for the checklist a live run must satisfy; firewall-aiops doctor (a firmware/version query on both platforms) is the fastest connectivity check. Both platforms are free and self-hostable (OPNsense is fully open-source; pfSense CE is free), so a home lab is the easiest place to run it.

  • Missing a capability? Open an issue or PR at github.com/AIops-tools/Firewall-AIops — contributions and feedback welcome.

License

MIT — see LICENSE.

Available Tools

35 tools
add_alias_entryA
Destructive

[WRITE][risk=medium] Add one entry to an alias; reversible (undo removes it).

Captures the alias' prior entries before the change. Pass dry_run=True to preview.

Args: name: Alias name (from list_aliases). entry: The host/network/port to add. dry_run: If True, preview without changing. target: Firewall target name from config; omit for the default.

Inputs: name (required), entry (required), target, dry_run.

alias_entriesA
Read-only

[READ] The member entries (hosts/networks/ports) of one alias.

Args: name: Alias name (from list_aliases). target: Firewall target name from config; omit for the default.

Inputs: name (required), target.

apply_changesA
Destructive

[WRITE][risk=high] Commit staged firewall config — makes edits live.

This is the "make it live" step after staged edits (e.g. toggle_rule).

Reads the staged rule set first and REFUSES when committing it would provably cut the endpoint this tool manages the firewall through — disabling the 'pass' rule that permits management access, or enabling a 'block' rule that covers it, locks out this tool and the undo that would reverse it. Uncertain cases (alias destinations, 'any', interface groups) warn and proceed.

dry_run=True returns the full staged change set and its assessment, not just an acknowledgement — use it before every apply.

Args: dry_run: If True, return the staged change set + lockout assessment without applying. override: Proceed despite a certain lockout finding. Only for operators with console / out-of-band access who mean it. target: Firewall target name from config; omit for the default.

Inputs: target, dry_run, override.

blocked_traffic_rcaA
Read-only

[READ] Rank the noisiest blocked sources and classify cause + action.

Keeps only blocked firewall-log entries, aggregates by source (hit count, distinct destination ports, busiest port), ranks the noisiest sources, and classifies each as a port scan, a service probe/brute-force on a sensitive port, or generic — with a recommended action. Every entry carries its numbers. Pass 'log_entries' for pure analysis, or a target to pull live.

Args: top: How many source rows to return, noisiest first (default 20). limit: How many recent log rows to pull when live (default 500). log_entries: Injected rows {action, source, destination, destinationPort, protocol}; skips the live pull. target: Firewall target name from config; omit for the default.

Returns dict: {blocksEvaluated, distinctSources, topSources:[{source, hits, distinctPorts, topPort, topPortHits, cause, action}], note}.

Inputs: top, limit, target, log_entries.

dhcp_leasesA
Read-only

[READ] Active DHCP leases (IP, MAC, hostname, state).

Args: online_only: If True, return only leases marked online; leases whose state the platform did not report are counted in onlineUnknown. target: Firewall target name from config; omit for the default.

Inputs: target, online_only.

dhcp_static_mappingsB
Read-only

[READ] DHCP static (reserved) mappings (MAC ↔ IP).

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

firewall_logA
Read-only

[READ] Recent firewall-log entries, optionally filtered to pass/block.

Args: action: Optional filter — one of pass, block, reject, rdr, nat. limit: Max entries to return (default 200). target: Firewall target name from config; omit for the default.

Returns an envelope: {"entries": [...], "returned": N, "limit": L, "truncated": bool}. When "truncated" is true there are more entries than were returned — re-run with a higher limit rather than treating the result as the complete log.

Inputs: limit, action, target.

firmware_statusA
Read-only

[READ] Firmware / OS version and whether updates are available.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

gateway_health_rcaA
Read-only

[READ] Rank gateways by loss + latency, map each down/degraded one to cause + action.

The flagship WAN RCA: pulls gateway status, flags each gateway that is down (status down/offline or 100% loss) or degraded (loss/latency over threshold), ranks worst-first, and attaches a likely cause and a recommended action. Every ranking carries its numbers, not a black-box verdict. Pass 'gateways' for pure analysis, or a target to pull live.

Args: loss_pct: Loss %% at/above which a gateway is degraded (default 5.0). latency_ms: RTT ms at/above which a gateway is degraded (default 150). gateways: Injected rows {name, address, status, lossPercent, rttMs}; skips the live pull. target: Firewall target name from config; omit for the default.

Returns dict: {gatewaysEvaluated, downCount, degradedCount, thresholds, worst:[{name, address, status, lossPercent, rttMs, down, degraded, cause, action}], note}.

Inputs: target, gateways, loss_pct, latency_ms.

gateway_statusB
Read-only

[READ] WAN/LAN gateways with status, loss %, and RTT latency.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

health_statusA
Read-only

[READ] System health snapshot: hostname, uptime, CPU, memory, load.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

interface_statusA
Read-only

[READ] Interfaces with link status + address, down interfaces first.

Use to identify failed or unaddressed interfaces on the selected firewall; use gateway_status for upstream reachability and packet-loss checks. Returns total and interfaces or an explicit error.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

ipsec_sasA
Read-only

[READ] IPsec security associations (phase-1/phase-2) with state.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

kill_statesA
Destructive

[WRITE][risk=medium] Flush pf state-table entries (optionally one source IP).

Drops tracked connections; they re-establish on the next packet. This includes THIS tool's own connection state, so the call may appear to fail even though the flush ran — that is a lost response, not a lockout, and a blind retry is the wrong reaction. Access is not lost: the permitting rule is untouched. Pass dry_run=True to preview.

Args: filter_ip: Optional source IP to scope the flush to (blank = all states). dry_run: If True, preview without flushing. target: Firewall target name from config; omit for the default.

Inputs: target, dry_run, filter_ip.

list_aliasesA
Read-only

[READ] All firewall aliases (name, type, description, member count).

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

list_rulesA
Read-only

[READ] List filter rules (optionally on one interface), normalized.

Args: interface: Optional interface name to filter by (e.g. wan, lan). target: Firewall target name from config; omit for the default.

Inputs: target, interface.

nat_one_to_oneB
Read-only

[READ] 1:1 NAT mappings (external ↔ internal), normalized.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

nat_outboundB
Read-only

[READ] Outbound (source) NAT mappings, normalized.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

nat_port_forwardsA
Read-only

[READ] Inbound port-forward (DNAT) rules, normalized.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

openvpn_sessionsB
Read-only

[READ] OpenVPN sessions / connected clients (name, address, bytes).

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

pending_changesA
Read-only

[READ] The staged rule set apply_changes would commit, with lockout risk.

Run this BEFORE apply_changes. It reports whether committing the staged rules would cut the endpoint this tool manages the firewall through — a disabled 'pass' rule that permits management access, or an enabled 'block' rule that covers it. Findings are ranked worst-first and carry a certain flag: certain ones make apply_changes refuse, uncertain ones (alias destinations, 'any', interface groups) are warnings only and never block.

basis states what this is: the staged rule STATE, not a diff against the running config — neither platform exposes a per-rule dirty flag over REST.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

rebootA
Destructive

[WRITE][risk=high] Reboot the firewall. IRREVERSIBLE — audit only, no undo.

Pass dry_run=True to preview.

Args: dry_run: If True, preview without rebooting. target: Firewall target name from config; omit for the default.

Inputs: target, dry_run.

reconfigureA
Destructive

[WRITE][risk=high] Reload/commit a subsystem's config (filter/nat/aliases).

Reloading the 'filter' subsystem commits the staged rule set just as apply_changes does, so it carries the same lockout guard and the same override.

Args: subsystem: Config subsystem to reload (filter, nat, aliases). dry_run: If True, preview without reconfiguring. override: Proceed despite a certain lockout finding (filter subsystem). target: Firewall target name from config; omit for the default.

Inputs: target, dry_run, override, subsystem.

remove_alias_entryA
Destructive

[WRITE][risk=medium] Remove one entry from an alias; reversible (undo adds it back).

Captures the alias' prior entries before the change. Pass dry_run=True to preview.

Args: name: Alias name (from list_aliases). entry: The host/network/port to remove. dry_run: If True, preview without changing. target: Firewall target name from config; omit for the default.

Inputs: name (required), entry (required), target, dry_run.

restart_serviceA
Destructive

[WRITE][risk=medium] Restart a firewall service (unbound, dhcpd, openvpn, ...).

REFUSES the service that answers this appliance's own management API (nginx / lighttpd / configd / webgui and friends): restarting it kills the connection this tool is using, so the restart cannot be observed and the undo cannot run. Restart those from the console instead. Pass dry_run=True to preview.

Args: service: Service name to restart. dry_run: If True, preview without restarting. target: Firewall target name from config; omit for the default.

Inputs: service (required), target, dry_run.

rule_detailA
Read-only

[READ] One rule's full detail by uuid/id.

Args: uuid: Rule uuid (OPNsense) or id (pfSense), from list_rules. target: Firewall target name from config; omit for the default.

Inputs: uuid (required), target.

rule_hit_and_shadow_analysisA
Read-only

[READ] Flag never-hit enabled rules and shadowed/redundant rules.

Finds enabled rules with zero evaluations (dead or misordered), rules fully covered by an earlier terminating rule (shadowed), and rules identical to an earlier one (redundant). Rules are compared in list order, top-down, exactly as pf evaluates them; every finding names the offending/covering rule uuid. Pass 'rules' for pure analysis, or a target to pull live.

Args: rules: Injected rows {uuid, sequence, enabled, action, interface, protocol, source, destination, destinationPort, evaluations}. interface: Optional interface filter when pulling live. target: Firewall target name from config; omit for the default.

Returns dict: {rulesEvaluated, unusedCount, shadowedCount, redundantCount, unusedRules, shadowedRules, redundantRules, note}.

Inputs: rules, target, interface.

rule_statesB
Read-only

[READ] Active pf state-table entries associated with rules (top-N).

Args: top: How many state entries to return (default 50). target: Firewall target name from config; omit for the default.

Inputs: top, target.

rule_statsB
Read-only

[READ] Per-rule hit counts / evaluations, busiest first (top-N).

Args: top: How many rules to return, busiest first (default 20). target: Firewall target name from config; omit for the default.

Inputs: top, target.

states_tableA
Read-only

[READ] Active pf state-table entries (connections currently tracked).

Args: top: How many state entries to return (default 100). target: Firewall target name from config; omit for the default.

Returns an envelope: {"states": [...], "returned": N, "limit": L, "truncated": bool, "total": T}. "truncated" is true when the state table holds more entries than were returned.

Inputs: top, target.

toggle_ruleA
Destructive

[WRITE][risk=medium] Enable/disable a filter rule; reversible.

Reads the rule first so the harness records an undo that restores its prior enabled state. Staged only — run apply_changes to make it live. Pass dry_run=True to preview.

Args: uuid: Rule uuid (OPNsense) or id (pfSense), from list_rules. enable: True to enable the rule, False to disable. dry_run: If True, preview without changing. target: Firewall target name from config; omit for the default.

Inputs: uuid (required), enable (required), target, dry_run.

top_talkersA
Read-only

[READ] Busiest source hosts, aggregated from the state table by bytes.

Args: top: How many talkers to return (default 20). target: Firewall target name from config; omit for the default.

Returns an envelope: {"topTalkers": [...], "returned": N, "limit": L, "truncated": bool, "total": T}. "truncated" is true when more distinct sources were seen than were returned.

Inputs: top, target.

undo_applyA
Destructive

[WRITE][risk=medium] Apply a recorded undo by dispatching its inverse tool.

The inverse runs through its own governed tool, so it is audited and labelled with its own risk tier there. Pass dry_run=True to preview the inverse call without executing it. A token can only be applied once.

Args: undo_id: The undoId from undo_list (or an _undo_id in a write result). dry_run: If True, preview the inverse tool + params without running it. target: Passed through to the inverse tool when it accepts a target.

Inputs: undo_id (required), target, dry_run.

undo_listA
Read-only

[READ] List recorded, not-yet-applied undo tokens (most recent first).

Each entry names the original tool, the inverse tool that undo_apply would run, and a human note. Use the undoId with undo_apply.

Returns {"undos": [...], "returned": N, "limit": L, "truncated": bool}. truncated is measured (one extra row is fetched), not guessed from a length coincidence: when it is true there are MORE tokens than shown, so re-run with a higher limit rather than reporting the list as complete.

Each entry carries effectVerified. False means the original write lost its response, so the change it reverses is PROBABLE, not confirmed — check the live state before applying, and do not report the result as a restore of a state that may never have been reached.

Args: limit: Max rows to return (default 50, capped at 500). target: Unused (undo state is host-local); accepted for CLI uniformity.

Inputs: limit, target.

wireguard_statusB
Read-only

[READ] WireGuard peers with connected state, last handshake, transfer.

Args: target: Firewall target name from config; omit for the default.

Inputs: target.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.6/5.0

    Scored across 35 tools

    Disambiguation4/5

    Most tools are clearly separated by resource and action, with read/write tags and cross-references that help distinguish similar areas (interface_status vs gateway_status, list_rules vs rule_detail). Some overlaps remain: states_table, rule_states, and top_talkers all expose pf state data, while apply_changes and reconfigure(filter) both commit staged rules. Overall the set is mostly distinct, but a few boundaries require careful reading.

    Naming Consistency4/5

    All tool names use snake_case and are generally readable. Mutating operations mostly follow verb_noun patterns such as toggle_rule, add_alias_entry, and remove_alias_entry, but many read/status tools use noun_status or noun phrases like interface_status and gateway_status, with a few bare verbs like reboot and reconfigure. The pattern is mostly consistent with minor deviations, not chaotic.

    Tool Count2/5

    At 35 tools the surface is comprehensive but heavy, exceeding the 25+ threshold where the set starts to feel over-scoped. Multiple status and RCA tools could be consolidated, and the broad domain only partly justifies the count. It is useful coverage but not a tightly scoped tool set.

    Completeness3/5

    Read coverage is strong across interfaces, gateways, rules, NAT, DHCP, VPNs, logs, health, firmware, and RCA workflows, with undo and staged apply support. However the write surface is thin: there is no create/edit/delete for firewall rules or aliases, and NAT, DHCP static mappings, and VPNs are read-only. That leaves notable lifecycle gaps for a firewall operations server.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Governed MySQL + MariaDB DBA operations — slow-query, lock-wait/deadlock, replication, and fragmentation RCA, with unbypassable audit logging (MCP + CLI), budget/runaway guards, dry-run, and undo/rollback.
      35
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Governed Prometheus + Grafana operations — firing-alert and scrape-target RCA, alert noise/flapping analysis, silences, and dashboards, with unbypassable audit logging (MCP + CLI), budget/runaway guards, dry-run, and undo/rollback.
      39
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Governed XCP-ng operations via Xen Orchestra REST — VM-health, SR-usage, backup-failure, and patch-state RCA, with unbypassable audit logging (MCP + CLI), budget/runaway guards, dry-run, and undo/rollback.
      29
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Governed SSO/IAM operations for Keycloak and Authentik — login-failure, stale-permission, client-config, and MFA RCA, with unbypassable audit logging (MCP + CLI), budget/runaway guards, dry-run, and undo/rollback.
      29
      MIT