A Model Context Protocol (MCP) server for managing Google Cloud NetApp Volumes (GCNV) resources through AI assistants such as Gemini CLI, Cursor, and other MCP-compatible clients.

• Node.js 20.19.0 or higher
• A Google Cloud project with the NetApp Volumes API enabled
• Google Cloud authentication credentials (see Authentication)

Run the published package directly (no local build required):

npx gcnv-mcp-server@latest --transport stdio

Most MCP clients need the same stdio server command:

{
  "command": "npx",
  "args": ["-y", "gcnv-mcp-server@latest", "--transport", "stdio"]
}

For clients that use the common mcpServers JSON shape:

{
  "mcpServers": {
    "gcnv": {
      "command": "npx",
      "args": ["-y", "gcnv-mcp-server@latest", "--transport", "stdio"]
    }
  }
}

For VS Code-style MCP configuration:

{
  "servers": {
    "gcnv": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "gcnv-mcp-server@latest", "--transport", "stdio"]
    }
  }
}

Client config file locations and JSON wrappers vary by application. See the official MCP server setup docs and the VS Code MCP configuration reference for client-specific details.

Claude Code:

claude mcp add gcnv -- npx -y gcnv-mcp-server@latest --transport stdio
claude mcp list

Codex CLI:

codex mcp add gcnv -- npx -y gcnv-mcp-server@latest --transport stdio
codex mcp list

See the Claude Code MCP quickstart and OpenAI Codex MCP docs for more client-specific options.

Gemini CLI extension users can install this repository as an extension:

# 1. Authenticate
gcloud auth login
gcloud auth application-default login

# 2. Install the extension
gemini extension install <repository-url>

# 3. Verify
gemini mcp list

Gemini automatically starts the MCP server when a linked extension needs it. No manual npm start is required for normal usage.

Antigravity / AGY users can install the repository as a plugin:

agy plugin install <repository-url>
agy plugin list

If you previously installed the Gemini CLI extension, migrate existing Gemini extensions:

agy plugin import gemini

You can also configure the MCP server directly in the shared MCP config at ~/.gemini/config/mcp_config.json:

{
  "mcpServers": {
    "gcnv": {
      "command": "npx",
      "args": ["-y", "gcnv-mcp-server@latest", "--transport", "stdio"]
    }
  }
}

Restart Antigravity or run agy, then use /mcp to verify the server is loaded. See the Google Antigravity MCP codelab for more details on Antigravity MCP configuration.

This repository includes assistant context files for clients that support repository guidance:

• AGENTS.md — shared GCNV safety and operating guidance.
• CLAUDE.md — Claude/Cursor compatibility shim that points to AGENTS.md.
• GEMINI.md — Gemini CLI extension context.

These files do not require a separate install step. Compatible clients load them automatically when using the repo or extension.

Google NetApp Volumes API calls need a Google access token. How that token is supplied depends on the MCP transport and caller.

Use ADC when running locally with Gemini CLI, Cursor, Claude Code, or other stdio MCP clients:

gcloud auth application-default login

Or point at a service account key:

export GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json

No extra environment variables are required. The server obtains a token from ADC for each tool call.

For HTTP or SSE transport, send a bearer token on each MCP HTTP request to override ADC for that request only. This path is independent of stdio delegated auth.

By default the server reads Authorization: Bearer <token>. Customize the header name with GCNV_AUTH_HEADER:

export GCNV_AUTH_HEADER="X-Google-Access-Token"

Then send requests with:

X-Google-Access-Token: Bearer <token>

Stdio MCP has no HTTP headers. A trusted parent process that spawns this server as a subprocess can inject a per-user Google access token via a runtime-only tool argument: _stdio_delegated_google_access_token.

This is disabled by default so public stdio clients cannot supply arbitrary bearer tokens via tool arguments.

Enable only when the MCP server is spawned by a fully controlled trusted subprocess:

export GCNV_STDIO_DELEGATED_ACCESS_TOKEN=true

When enabled:

• The runtime arg is never advertised in list_tools schemas.
• It is stripped before tool handlers run and is never logged.
• It overrides ADC for that tool call only.
• HTTP clients should continue to use the bearer header path above; Gemini CLI and other public stdio clients should use ADC.

To externalize ONTAP discover ranking, set a knowledge endpoint and the server will call it per query.

Request/response contract is documented in docs/ontap-kg-protocol.md and schema in schemas/ontap-kg.schema.json.

When the KG request fails or returns invalid payload, discover falls back to the bundled ontap-api-index.json.

The server supports stdio (default) and HTTP/SSE transports.

Used automatically by Gemini CLI and other stdio-based MCP clients.

npm start                          # default stdio
npm run start:stdio                # explicit

For web-based MCP clients or remote access.

npm run start:http                 # port 3000
npm start -- -t http -p 8080       # custom port

HTTP endpoint: http://localhost:<port>/message

Security note: The HTTP/SSE transport does not provide built-in TLS. If you expose it beyond a local trusted client, use customer-managed TLS and network access controls, such as a reverse proxy, VPN, service mesh, or SSH tunnel. Do not expose the MCP HTTP endpoint directly to the public internet.

Service level guidance:

• FLEX -- Smaller minimums, broader region availability, independent performance scaling. Minimum: 1024 GiB (FILE/UNIFIED) or 6144 GiB (UNIFIED large capacity).
• STANDARD / PREMIUM / EXTREME -- Classic tiers with fixed performance-to-capacity ratio. Minimum: 2048 GiB.
• serviceLevel is accepted case-insensitively (e.g. flex or FLEX).
• FLEX pools in a region-level location require both zone and replicaZone; zone-level locations satisfy this automatically.
• storagePoolType accepts FILE or UNIFIED; UNIFIED is only available for FLEX.
• scaleType: only set to SCALE_TYPE_SCALEOUT when creating a large capacity FLEX UNIFIED pool. Omit for all other pools (defaults to SCALE_TYPE_DEFAULT).
• mode: defaults to DEFAULT for a standard pool, or set to ONTAP to create an ONTAP-mode pool that exposes advanced ONTAP features (SnapMirror, QoS policies, direct ONTAP REST access) via the ONTAP Expert Mode Tools. ONTAP mode requires serviceLevel: FLEX and storagePoolType: UNIFIED.

iSCSI notes: Protocols must be ["ISCSI"] only (no mixing). Requires hostGroup or hostGroups. Optional blockDevice object with identifier, osType (LINUX / WINDOWS / ESXI), and sizeGib.

Large capacity volumes: Some workloads require larger volumes and higher throughput, which can be achieved by using the large capacity volume option for these service levels. Large capacity volumes provide six storage endpoints (IP addresses) to load-balance client traffic to the volume and deliver higher performance.

• FLEX Unified: Large capacity volumes can be sized between 4.8 TiB and 2.48 PiB, or up to 20 PiB with auto-tiering, in increments of 1 GiB, and deliver throughput performance of up to 22 GiBps. The FLEX storage pool must be created with scaleType: SCALE_TYPE_SCALEOUT; non-scale-out FLEX pools cannot host large capacity volumes. When largeCapacityConstituentCount is explicitly set, the minimum drops to 2.4 TiB (2,400 GiB) — FLEX Unified only. When largeCapacityConstituentCount is omitted, the 4.8 TiB (4,916 GiB) floor applies.
• Premium and Extreme: Large capacity volumes can be sized between 15 TiB and 3 PiB in increments of 1 GiB, and deliver throughput performance of up to 30 GiBps.

A large capacity volume is internally composed of several constituent volumes (FlexVols) distributed across the pool's storage aggregates. On gcnv_volume_create you can optionally pass largeCapacityConstituentCount to control how many constituents the volume is built from:

• Minimum: 2.
• Default (when omitted): chosen by the backend based on the active deployment layout.
• The constituent count must be chosen at create time and cannot be changed later.

SMB attributes: When protocols includes SMB, gcnv_volume_create accepts optional SMB feature flags that map to the smbSettings field on the volume:

• smbEncryptData: true → ENCRYPT_DATA (require SMB encryption in flight)
• smbHideShare: true → NON_BROWSABLE (hide the share from browse lists)
• smbAccessBasedEnumeration: true → ACCESS_BASED_ENUMERATION (ABE) — controls the visibility of files and folders based on the permissions assigned to the user
• smbContinuouslyAvailable: true → CONTINUOUSLY_AVAILABLE (CA share for SQL Server / FSLogix; choice is permanent on the volume)
• smbSettings: ["OPLOCKS", ...] — additional API enum values; merged with the booleans above. Do not pass SMB_SETTINGS_UNSPECIFIED; BROWSABLE and NON_BROWSABLE (or BROWSABLE together with smbHideShare) are rejected. NON_BROWSABLE and CONTINUOUSLY_AVAILABLE together (or smbHideShare with smbContinuouslyAvailable) are also rejected — CA shares must be browsable.

These flags require protocols to include SMB. CONTINUOUSLY_AVAILABLE is not supported on FLEX storage pools — the request is rejected before reaching the API. Use STANDARD, PREMIUM, or EXTREME for CA shares.

FlexCache: On default-mode pools, use gcnv_volume_create with cacheParameters to create a cache volume (origin cluster, SVM, volume, and intercluster LIF IPs). Use gcnv_volume_update with cacheParameters.cacheConfig to change cache settings. On ONTAP-mode pools, use ONTAP Expert Mode Tools instead.

Replication is supported between specific region pairs (Standard/Premium/Extreme) or within the same region group (Flex). See the replication guide.

ONTAP Expert Mode is available for storage pools created with mode: ONTAP. It exposes supported ONTAP REST operations through the MCP server while still using Google Cloud authentication and the GCNV control plane proxy. No separate ONTAP credentials or direct ONTAP endpoint configuration are required.

For advanced resources such as QoS policies, SnapMirror, export policies, CIFS services, igroups, snapshot policies, SnapLock, Event Based Retention, schedules, and cluster/SVM peering, use ontap_discover first and then call ontap_execute with the discovered method, path, and body template.

ontap_discover responses include descriptions, hints, example body templates, and generated requiredBody metadata when ONTAP swagger marks body fields as required. ontap_execute preflight-validates requests against the bundled API index (method, path, and required body fields) before calling ONTAP. Unlisted paths — including GET — are rejected with a scope_denied envelope; use ontap_discover to find supported endpoints. For POST/PATCH calls, pass the body as a JSON string.

ONTAP mutating operations commonly return async jobs. After create, update, or replication actions, poll the returned job UUID with ontap_job_get; a returned job reference does not by itself mean the operation is complete.

Safety guardrails:

• Delete operations are not supported. This server exposes no delete tools, and ontap_execute rejects the DELETE method. Delete resources through the Google Cloud console, gcloud, or another tool.
• Some endpoints are out of scope for this MCP server or blocked by the proxy/RBAC policy. These return a scope_denied envelope and should not be retried with sibling paths or private CLI variants.
• /api/private/cli subpaths are intentionally blocked at preflight. Use public /api/... endpoints from ontap_discover or dedicated tools instead.
• User-facing clients should redact sensitive fields before displaying tool output.

src/
  index.ts                          # Entry point (stdio + HTTP/SSE transports)
  logger.ts                         # Structured logging (pino)
  registry/
    register-tools.ts               # Tool registration
  resources/
    ontap-api-index.json            # Static ONTAP REST API catalog (powers ontap_discover)
  tools/
    *-tools.ts                      # Tool definitions (Zod schemas)
    ontap-*-tool.ts                 # ONTAP Expert Mode tool schemas
    handlers/
      *-handler.ts                  # Tool implementations
  types/
    tool.ts                         # Shared TypeScript interfaces
  utils/
    netapp-client-factory.ts        # NetApp client factory with caching
    ontap-http-client.ts            # ONTAP REST client (auth, body envelope, response unwrapping)
    ontap-index-loader.ts           # Loads and indexes the ONTAP API catalog
    ontap-preflight-validator.ts    # Pre-execution validation (index allowlist, CLI block, required body)
    ontap-response-utils.ts         # ONTAP response shaping and error helpers
    ontap-audit-logger.ts           # Optional Markdown audit log for ONTAP tool calls
    scope-denied-envelope.ts        # Proxy/RBAC scope denial envelope

ontap_execute returns scope_denied: ONTAP API index could not be loaded ... — the bundled API index file is missing or unreadable, usually from an incomplete install. Re-add gcnv-mcp-server in your MCP client to trigger a fresh npx download; if that doesn't help, run npm cache clean --force and re-add. Persistent failures: file an issue at https://github.com/NetApp/gcnv-mcp-server with the server log.

npm install
npm run build          # lint + format + compile
npm test               # run all tests
npm run test:coverage  # with coverage report
npm run sbom           # print an SPDX SBOM

npm run dev            # stdio via tsx
npm run dev:http       # HTTP via tsx

1. Define the tool schema in src/tools/<resource>-tools.ts
2. Implement the handler in src/tools/handlers/<resource>-handler.ts
3. Register the tool in src/registry/register-tools.ts

npm run githooks:install   # enables lint + test on commit

Pricing is based on provisioned pool capacity, not consumed capacity. Some features (e.g. auto-tiering) add usage-based I/O charges. See the pricing page or use the Google Cloud Pricing Calculator for estimates.

Apache-2.0

We'd love to hear from you. Share feedback, feature requests, or bug reports at ng-gcnv-mcp-feedback@netapp.com.

Contributions are welcome. Please open an issue or submit a pull request.