Both installation paths make surface area a decision rather than a default: locally the --services flag loads only the services you name, and remotely each service is its own endpoint you add or leave out. The project's own advice is to enable only what you need, because every extra service is context the model carries on every turn.
A Go server built on DigitalOcean's own godo library that exposes the DigitalOcean API as MCP tools. It comes two ways: a hosted HTTPS endpoint per service that you connect to without running anything, or a local process where a --services flag decides which services load.
- App Platform end to end: apps-list and apps-get-info to see what is running, apps-create-app-from-spec to deploy from a Git repo, Dockerfile or container image, apps-update to change a live spec, apps-delete to tear one down
- apps-get-deployment-status returns the active deployment and its health, and apps-get-logs hands back both live and historic log URLs for a specific component
- region-list, so a request for capacity is answered against the regions and droplet sizes that actually exist rather than a guess
- Droplets managed as a fleet — droplet-resize, droplet-enable-backups and the rest of the droplet surface, through the droplets service
- Managed databases, Kubernetes clusters and the container registry each behind their own service: Postgres, MySQL and Redis clusters, DOKS node pools, DOCR repositories, tags, manifests and garbage collection
- Networking as one surface: domain-create, DNS records, certificates, firewalls, load balancers, reserved IPs, BYOIP prefixes, VPCs, cdn-flush-cache and vpc-peering-create
- The GenAI side split by job — batch inference, custom models, inference routers, dedicated GPU inference, evaluation workflows, and inference-model-catalog-search plus inference-model-catalog-get-card for choosing a model
- Storage covered by three services: Spaces with its S3-compatible access keys, block storage volumes and snapshots, and NFS file shares
- docs-search, docs-get-quickstart, docs-troubleshoot and docs-get-related read DigitalOcean's public documentation — that endpoint is the one service that needs no API token
- Insights watches resources and endpoints and raises an alert when something is slow, down, or holding a certificate about to expire
A DigitalOcean API token, scoped when you generate it, held in DIGITALOCEAN_API_TOKEN and sent as an Authorization: Bearer header to a hosted endpoint such as https://apps.mcp.digitalocean.com/mcp. Running it locally instead needs Node.js v18 or later and NPM v8 or later, and takes the shape npx @digitalocean/mcp --services apps,droplets. The README is firm that the token belongs in an environment variable, never pasted into a config file that might reach GitHub.
One command plus a key — npx @digitalocean/mcp --services apps, then supply credentials