Skip to content

Installation

Harvest MCP Server is distributed as a Docker container image, or you can build it from source.

Container Images

Harvest MCP Server is available as pre-built container images:

Image Description
ghcr.io/netapp/harvest-mcp:latest Stable release version
ghcr.io/netapp/harvest-mcp:nightly Latest development builds

MCP Client Integration

Stdio Mode

For MCP clients like GitHub Copilot, add to your mcp.json:

{
  "servers": {
    "harvest-mcp": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "--env", "HARVEST_TSDB_URL=http://your-prometheus:9090",
        "ghcr.io/netapp/harvest-mcp:latest"
      ]
    }
  }
}

HTTP Mode

For HTTP-based MCP clients, first start the server:

docker run -d \
  --name harvest-mcp-server \
  -p 8082:8082 \
  --env HARVEST_TSDB_URL=http://your-prometheus:9090 \
  ghcr.io/netapp/harvest-mcp:latest \
  start --http --port 8082 --host 0.0.0.0

If you only want to bind to localhost, omit the --host option.

Then configure your mcp.json:

{
  "servers": {
    "harvest-mcp": {
      "type": "http",
      "url": "http://localhost:8082"
    }
  }
}

For remote server access:

{
  "servers": {
    "harvest-mcp": {
      "type": "http",
      "url": "http://your-server-ip:8082"
    }
  }
}

Building from Source

Prerequisites

  • Go(check .harvest.env in the repository root for the exact required version)
  • Git
  • Make
  • Docker (optional, for building Docker images)

Clone the Repository

First, clone the Harvest repository:

git clone https://github.com/NetApp/harvest.git
cd harvest

Build Docker Image

Build your own Docker image from source:

# Navigate to the mcp directory
cd mcp

# Build the Docker image using make (creates harvest-mcp:latest by default)
make docker-build

# Or specify a custom tag
make docker-build DOCKER_TAG=harvest-mcp:local

Alternatively, build directly with Docker:

# From the harvest repository root
docker build -f mcp/Dockerfile -t harvest-mcp:local .

Running the Built Docker Image

After building, use your local image in your MCP client configuration. See MCP Client Integration above for configuration examples - just replace ghcr.io/netapp/harvest-mcp:latest with your local image tag (e.g., harvest-mcp:local).

Build Native Binary

Create a standalone binary package:

# Navigate to the mcp directory
cd mcp

# build for specific platforms:
GOOS=linux GOARCH=amd64 make package    # Linux AMD64
GOOS=darwin GOARCH=arm64 make package   # macOS ARM64 (Apple Silicon)

# Creates: dist/harvest-mcp-<version>-<release>_<os>_<arch>.tar.gz

Running the Native Binary

After extracting the package, configure it in mcp.json for MCP clients.

Configure in mcp.json

For MCP clients like GitHub Copilot, add to your mcp.json:

{
  "servers": {
    "harvest-mcp": {
      "type": "stdio",
      "command": "/path/to/harvest-mcp/bin/harvest-mcp",
      "args": ["start"],
      "env": {
        "HARVEST_TSDB_URL": "http://your-prometheus:9090"
      }
    }
  }
}

For HTTP mode, start the server first:

# Start the server
HARVEST_TSDB_URL=http://your-prometheus:9090 /path/to/harvest-mcp/bin/harvest-mcp start --http --port 8082

Then configure your mcp.json:

{
  "servers": {
    "harvest-mcp": {
      "type": "http",
      "url": "http://localhost:8082"
    }
  }
}

Logs

To view the MCP server logs:

docker logs <container-id>

Configuration

For complete configuration options and environment variables, run:

docker run --rm ghcr.io/netapp/harvest-mcp:latest start --help

This displays all available environment variables with descriptions, authentication options, and advanced settings.

Scoping to Harvest Metrics (Multi-source TSDB)

If your Prometheus or VictoriaMetrics instance contains metrics from multiple sources — for example, both Harvest and node_exporter, kube-state-metrics, or other exporters — use HARVEST_TSDB_LABEL_FILTER to scope every MCP query exclusively to Harvest metrics.

Environment variable: HARVEST_TSDB_LABEL_FILTER

Value: any valid PromQL label matcher, without the surrounding {}

Scenario Value
Job label job="harvest"
Custom label source="harvest"

When set, the filter is automatically injected into every query and match[] selector issued by the MCP server. You do not need to include it in any question you ask the AI — it is applied transparently.

Query type Before After (with job="harvest")
Bare metric volume_size_used_percent volume_size_used_percent{job="harvest"}
With existing labels volume_size_used_percent{svm="vs1"} volume_size_used_percent{svm="vs1",job="harvest"}
Aggregation count by (cluster) (cluster_new_status) count(cluster_new_status{job="harvest"}) by(cluster)
Nested function sum(rate(http_requests_total[5m])) sum(rate(http_requests_total{job="harvest"}[5m]))

Next Steps