Harvest Speaks: Turning ONTAP Metrics Into Conversations
Storage • Observability • AI Agents
Harvest Speaks: Turning ONTAP Metrics Into Conversations
How an old open-source workhorse quietly became an AI-ready observability layer — and what that means for the way we work with storage data.
⏰ 6 min read
There's a particular kind of open-source tool that never quite gets the spotlight it deserves. It doesn't have a flashy launch event. It doesn't trend on social media. It just sits there, quietly doing its job, one poller cycle at a time — until one day you realise half your monitoring stack has been leaning on it for years without you thinking twice.
For a lot of ONTAP and StorageGRID admins, that tool is NetApp Harvest. And this year, it picked up a capability that's worth a proper look: the ability to talk directly to AI agents.
What Harvest Actually Is
At its core, Harvest is an open-metrics collector. It reaches into ONTAP clusters, StorageGRID grids, E-Series arrays, and even Cisco Nexus switches, pulls out performance, capacity, and hardware metrics, and routes them into a time-series database of your choice — typically Prometheus or VictoriaMetrics. Grafana then sits on top, turning that data into the dashboards most storage teams are already familiar with.
It's been around long enough to become the default answer whenever someone asks "how do I monitor my ONTAP environment without buying another platform?" It's free, it's actively maintained, and the dashboard library has grown steadily — the most recent major release brought E-Series support to general availability, added dashboards for license tracking and projected time-till-full on aggregates and volumes, and extended coverage to Google Cloud NetApp Volumes and newer Cisco switch models.
The Familiar Architecture
If you've deployed Harvest before, this diagram will look familiar. If you haven't, this is the shape of it:
Nothing about that pipeline changes. Harvest still collects, Prometheus still stores, Grafana still visualises. What's new is that second branch on the right — a direct line from your metrics store into an AI agent.
A Harvest-powered Grafana dashboard, showing cluster-level performance and capacity metrics.
The Pivot: From Dashboards to Dialogue
Dashboards are great when you know what you're looking for. They're less great at 2am when you're trying to figure out, in plain language, "why is this SVM slow" or "which volumes are about to run out of space" — and you don't want to click through six panels to find out.
That's the gap the newest addition to Harvest is aimed at: a Harvest MCP server, built on the Model Context Protocol. Rather than replacing Prometheus or Grafana, it sits on top of the same pipeline you already run, and exposes your metrics through a structured interface that AI agents can query safely and consistently — no PromQL required.
MCP-compatible clients — GitHub Copilot, Claude Desktop, and similar tools are the ones NetApp calls out explicitly — can connect to this server and translate natural-language questions into structured queries against your existing metrics store, then hand back the answer in plain language. Your dashboards and alerts don't go away; this just gives you a faster, more conversational way to do the exploratory digging that used to mean opening Grafana and hunting.
Leveraging an Existing AI Agent
Here's the part that matters most for presales conversations and day-to-day ops: if you're already running Harvest, you don't need a new platform to get this. You're extending a pipeline that's probably already in production.
The exact configuration syntax will depend on which AI agent you're using and how you've deployed the MCP server, so I'd point you to NetApp's own Harvest documentation for the precise setup steps rather than me reproducing them here. But structurally, connecting an MCP server to a client like Claude Desktop typically comes down to adding an entry to that client's MCP configuration — something along these lines, illustratively:
Once that connection is live, the AI agent has read access to your Harvest-collected metrics for the questions you ask it — nothing more. It's an additive layer, not a replacement for the governance and access controls you already have around Prometheus and Grafana.
What This Looks Like in Practice
A few examples of the kind of questions this setup is built to answer — the sort of thing that used to mean five minutes of dashboard-hunting:
What's the overall health of my infrastructure right now?
A summarised sweep across clusters, aggregates, and volumes, instead of tabbing between separate Grafana panels.
Which volumes are running out of space?
A direct answer pulled from capacity metrics, without needing to know which dashboard has the right panel.
Which systems are experiencing high latency?
A cross-cluster performance check that would otherwise mean comparing several dashboards side by side.
Why This Matters Beyond the Ops Desk
From where I sit in presales, this is a useful shift to understand — not because it's a new product to pitch, but because it changes the shape of a conversation I have often. Customers with mature ONTAP environments frequently already have Harvest running somewhere. The question used to be "do you need a new monitoring platform to get AI-driven insights out of your storage estate?" Increasingly, the honest answer is: maybe not — you might just need to point the AI agent you already use at the pipeline you already have.
That's a more interesting conversation than a rip-and-replace pitch. It's also a good illustration of a broader pattern worth watching: infrastructure tooling isn't just adding AI features on top, it's exposing itself as data that AI agents can reason over directly, through open standards like MCP rather than proprietary integrations.
A Few Things Worth Keeping in Mind
This is still an evolving area, so a couple of caveats seem worth flagging as you evaluate it. First, an MCP layer is only as good as the metrics feeding it — if your Harvest pollers aren't configured to collect the data you care about, no amount of natural-language querying will conjure it. Second, treat access to an MCP server the way you'd treat any other read path into production monitoring data — with the same network segmentation and access review you'd apply to Grafana or Prometheus itself.
Closing Thought
Harvest has spent years being the tool nobody talks about because it just works. It's a fitting twist that the thing pushing it back into the conversation is, quite literally, conversation. If you're already running it, this might be the smallest lift you'll find all year toward making your storage estate genuinely AI-accessible — no new platform, no rip-and-replace, just a new way of asking the questions you were already asking.
Are you using NetApp Harvest? Share your thoughts and comment.
Comments