Fortimanager Ops

FortiManager Operations — the manager plane

MCP Server

• Server: fortinet-mcp (NetClaw-authored, spec 080 / roadmap R3)
• Command: $FORTINET_MCP_CMD
• Transport: stdio
• Requires: FORTIMANAGER_HOST, FORTIMANAGER_API_TOKEN
• Mode: read-only by default; installs require two gates (below)

v2.0.0: this skill previously declared FORTIMANAGER_MCP_CMD pointing at jmpijll/fortimanager-mcp, which was never vendored, never registered and not installable — the skill was a claim with no server behind it. It is now backed by fortinet-mcp, and the command variable changed to FORTINET_MCP_CMD because one server serves all three Fortinet planes.

The distinction this skill exists to protect

FortiManager holds INTENT. It does not know what a device is actually doing.

A policy package and a FortiGate's running config legitimately diverge between installs. That gap is where drift, unauthorised change and failed installs live — use fgt_compare_with_manager to surface it rather than assuming they agree.

Tools (8 read-only + 2 write)

Every response carries its plane and scope

{ "plane": "manager", "scope": {"adom": "root", "package": "Corp"},
  "source": "...", "outcome": "ok", "data": {...}, "notes": [] }

outcome distinguishes results that look alike: ok, empty_result, plane_unreachable, auth_expired, auth_missing, scope_indeterminate, and the three separate write refusals. An expired session is auth_expired, never "no policies exist" — that would be a silent, plausible, wrong answer.

Workflow: policy package audit

1. fmg_list_adoms → pick the ADOM. A package name is unique only within one.
2. fmg_list_policy_packages → find the package and its install targets.
3. fmg_get_policy_package → ordered rules. Note position: shadowing is positional.
4. fmg_resolve_object on every group a rule references. A rule reported only by object name is not an audit — "allow GRP_CORP to GRP_DMZ" says nothing about which addresses that permits.
5. fmg_get_revisions → rollback context before proposing any change.
6. Feed the rules to fwrule-analyzer for overlap, shadowing and conflict analysis.

Workflow: is intent matching reality?

1. fmg_get_policy_package — the intent.
2. fgt_compare_with_manager (in fortigate-ops) — the divergence.
3. only_in_device entries are candidate out-of-band changes: someone edited the firewall directly. This is the single most valuable finding here.
4. only_in_manager usually means the package has not been installed since those rules were added — check fmg_list_devices sync status.

Writes: two gates, and neither substitutes for the other

fmg_install_package pushes policy to production firewalls — the highest blast-radius action available here.

Human approval and a ServiceNow change record are different gates. A CR does not imply a human said yes; a human saying yes does not imply change control approved it. Lab devices waive the CR gate only — never the approval gate — and a device that cannot be classified is treated as production.

Always run fmg_preview_install first. It shows what would change and needs no gate.

Integration with other skills

Important rules

• Treat package install as production change execution — baseline, verify, rollback.
• Always name the ADOM. A package without one is ambiguous.
• Resolve objects before drawing conclusions.
• Never present manager intent as observed device state.
• An empty result is not an error, and auth_expired is not "no data".