Fabric AIops

Disclaimer: Community-maintained open-source project. Not affiliated with, endorsed by, or sponsored by Cisco, Meraki, Arista, Ubiquiti, or any network-controller vendor. "Cisco", "Meraki", "Catalyst", "DNA Center", "Arista", "CloudVision", "Ubiquiti", "UniFi" and all product/trademark names belong to their respective owners. MIT licensed.

Governed AI-ops for network fabrics managed through a controller — the Cisco Meraki Dashboard API (the reference platform, full read + write), Cisco Catalyst Center (formerly DNA Center; read subset), Arista CloudVision Portal (CVP) (read subset), and UniFi Network (self-hosted controller or UniFi OS console; read subset + device restart) — with a built-in governance harness: unified audit log, policy engine, token/runaway budget guard, undo-token recording, and graduated-autonomy risk tiers. Multi-platform by construction: a registry keyed by platform maps every canonical operation onto each controller's REST API (path templates + response adapters), so adding a controller is a registry entry, never new ops/CLI/MCP surface. An operation a platform doesn't map returns a clear teaching error ("not supported on X yet — open an issue"), never a silent no-op. The test suite is mock-based; no platform has yet been exercised against a live controller — see docs/VERIFICATION.md.

What it does

Three flagship signature analyses, plus the guarded reads and writes around them:

• Uplink loss & latency RCA — pull MX WAN uplink loss + latency across an org, rank the worst uplinks by a composite of average loss and latency, and map each degraded uplink to a likely cause + recommended action. Every ranking carries its numbers, not a black-box verdict.
• Network health score — a composite 0-100 score per network from device online %, uplink health %, and an alert-severity penalty (weighted 0.5/0.3/0.2), with every component returned so the number is explainable.
• Config template drift — for networks bound to a config template, list the settings that have drifted from the template (expected vs actual).

What works

• CLI (fabric-aiops ...): init, overview, org, network, device, client, health, remediate, secret, doctor, mcp.
• MCP server (fabric-aiops mcp or fabric-aiops-mcp): 34 tools (25 read, 9 write), every one wrapped with the bundled @governed_tool harness.
• Encrypted credentials: the controller secret (Meraki API key / Catalyst Center username:password / CVP service-account token / UniFi API key) lives in an encrypted store ~/.fabric-aiops/secrets.enc (Fernet + scrypt) — never plaintext on disk. Unlock with a master password from FABRIC_AIOPS_MASTER_PASSWORD (MCP/CI) or an interactive prompt (CLI).
• Reversibility: mutating writes fetch the real before-state first and record a faithful inverse (update_device/update_network_vlan restore prior values; claim↔remove; bind↔unbind/rebind). Irreversible ops (reboot_device, blink_device_leds) record the prior state for audit but declare no undo.
• Safety: every state-changing CLI op supports --dry-run and requires double confirmation; every write MCP tool takes a dry_run preview.

Capability matrix (34 MCP tools)

network_health_score and config_template_drift are injected-only (they score data you already hold); uplink_loss_and_latency_rca accepts injected records for offline analysis or pulls live from a configured target. Device models carry a product-type prefix: MX appliance, MS switch, MR wireless AP, MV camera, MG cellular gateway.

Platform support matrix

One tool, four controller platforms. The ops/CLI/MCP surface is identical everywhere; each platform maps the canonical operations it supports and raises a teaching error for the rest ("not supported on <platform> yet — open an issue or PR").

Concept mapping: canonical organizations/networks are Catalyst Center sites, CVP containers, and UniFi sites (the canonical id is the site's short name — the /api/s/{site}/ path segment); all have one global tree, so the org scope does not filter their lists. On unifi, device-scoped calls (device get / switch_ports / reboot) fill the site from the target's default org_id — set it in config.yaml (or the init wizard). Writes are Meraki-only except UniFi device restart — Catalyst Center and CVP change models (task/configlet workflows) don't map cleanly onto these canonical writes, so each write fails fast with a teaching error before any controller call (never a silent no-op). CVP config-drift surfaces through device_inventory (complianceCode/complianceIndication per device) and network_alerts (events); configlet-content retrieval and deep pagination on catalyst/cvp/unifi are known deferrals.

Per-platform auth

fabric-aiops init walks through the platform choice and stores the right kind of secret; fabric-aiops doctor probes each target with the canonical top-of-hierarchy read (organizations / sites / containers / UniFi sites), exercising the full auth flow.

Security: read-only mode

This tool is meant to be handed to an AI agent, so its safety story is enforced by the server rather than requested in a prompt:

export FABRIC_READ_ONLY=1

With that set, the 9 write tools are never registered. An MCP client lists 25 tools instead of 34 — the writes are not hidden, not gated behind a flag, and not merely refused when called. They are absent from the session. A model cannot invoke a tool it was never offered, and cannot be argued into one.

That distinction is the whole point. A tool that exists but refuses still invites retry loops and "I'll describe the call instead" behaviour from smaller models, and it leaves a reviewer trusting a promise. An absent tool is a fact you can check: connect, list the tools, and see that the writes are not there.

Enforcement is two layers deep, so the switch cannot be sidestepped by changing entry point:

Read operations are unaffected, and every call is still audited to ~/.fabric-aiops/audit.db.

The read/write split is derived from each tool's declared risk_level, and a test asserts that this never disagrees with the [READ]/[WRITE] tag in the tool's own documentation — so a write can't quietly present itself as a read.

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.

Quick start

uv tool install fabric-aiops              # or: pipx install fabric-aiops
fabric-aiops init                         # wizard: choose platform (meraki/catalyst/cvp/unifi) + store the secret (encrypted)
fabric-aiops doctor                       # verify config, secrets, connectivity (per-platform auth probe)
fabric-aiops overview                     # one-shot fabric fleet health
fabric-aiops health uplink-rca            # rank worst MX WAN uplinks + cause/action (meraki)
fabric-aiops device inventory --model MS  # switches in the org

Run as an MCP server (stdio):

export FABRIC_AIOPS_MASTER_PASSWORD=...   # unlock secrets non-interactively
fabric-aiops-mcp

Governance

Every MCP tool passes through the bundled @governed_tool harness:

• Audit — every call (params, result, status, duration, risk tier, approver, rationale) is logged to ~/.fabric-aiops/audit.db (relocatable via FABRIC_AIOPS_HOME).
• Budget / runaway guard — token and call budgets trip a circuit breaker.
• Risk tiers — graduated autonomy; high-risk ops can require a named approver (FABRIC_AUDIT_APPROVED_BY / FABRIC_AUDIT_RATIONALE — the env-var names the bundled harness reads).
• Undo recording — reversible writes record an inverse descriptor built from the fetched before-state.

Scope

This is the network-fabric / controller member of the AIops-tools family (governed AI-ops with audit + budget + undo + risk tiers). Do NOT use it for OT / industrial edge (Modbus, OPC-UA, PROFINET) — see the separate industrial-aiops line — nor for device-level CLI/SSH network automation.

Missing a capability?

Coverage is intentionally a curated subset of each controller's API. Missing a call or a device family on Meraki? A ❌ in the support matrix you need on Catalyst Center, CloudVision Portal, or UniFi Network (writes included — e.g. the UniFi cookie-login fallback for pre-9.0 controllers)? Want another controller platform entirely? Open an issue or PR — contributions welcome (a platform is a single descriptor module: path templates + response adapters).

Status

The test suite is mock-based. No platform has yet been exercised against a live Meraki organization, Catalyst Center appliance, CloudVision Portal instance, or UniFi controller — all four platforms' API paths are modelled from the public API shapes. docs/VERIFICATION.md defines the checklist a live run must cover; fabric-aiops doctor is the fastest live check.