Automation Pilot MCP Server Generation

SAP Automation Pilot MCP Server Definition Generator

This skill guides the creation of MCP server definitions for SAP Automation Pilot. MCP server definitions are JSON files that expose Automation Pilot commands as MCP tools, allowing AI assistants to invoke them directly.

JSON Schema

Top-Level Structure

Every MCP server definition follows this structure:

{
  "name": "kebab-case-server-name",
  "enabled": true,
  "instructions": "<Natural-language description for the AI>",
  "mcpTools": [ ...tool objects... ]
}

Tool Object Structure

Each entry in mcpTools uses exactly this schema — all fields are required:

{
  "commandId": "<catalog>-<suffix>:<CommandName>:<version>",
  "name": "<snake_case_tool_name>",
  "enabled": true,
  "inputReferences": [
    "<catalog>-<<<TENANT_ID>>>:<InputName>:<version>"
  ],
  "tags": {},
  "title": "<Title Case Label>",
  "destructiveHint": false,
  "idempotentHint": false,
  "openWorldHint": true,
  "readOnlyHint": false
}

Command ID Format

Command IDs follow the pattern: <catalog>-<suffix>:<CommandName>:<version>

Suffix Rules

The <<<TENANT_ID>>> is a placeholder substituted at deployment time with the actual tenant identifier.

Known Catalog Prefixes

Input References (Credential Injection)

Input references inject pre-configured credentials into the command at runtime. Format: <catalog>-<<<TENANT_ID>>>:<InputName>:<version>

Known Credential Types

Rules

• Most tools have exactly one input reference for their credential type
• Use empty array [] when credentials are embedded directly in the command definition (e.g., custom commands with hardcoded config)
• A single server can mix credential types if tools span different services (e.g., App Logs & Metrics uses both AnsCredentials and BtpCredentials)
• The same credential reference can be reused across multiple tools in a server

Hint Boolean Rules

The four hint booleans form a semantic contract. Follow these rules strictly:

readOnlyHint

• true for all list_*, get_*, fetch_* operations that only read data
• false for all write, create, update, delete, start, stop operations
• If readOnlyHint: true then destructiveHint MUST be false

destructiveHint

• true ONLY for operations that delete or permanently alter state:
  • Delete operations (delete_*, remove_*)
  • Irreversible state changes (revert_to_snapshot, upgrade_instance)
• false for everything else (reads, creates, updates, start/stop)
• If destructiveHint: true then readOnlyHint MUST be false

idempotentHint

• true ONLY for write operations that are safe to repeat:
  • Idempotent updates (update_transport_node, update_transport_route)
  • Idempotent state transitions (start_instance, stop_instance, restart_instance, enable_capabilities)
• false for read operations (idempotent by nature but not tagged)
• false for creates, deletes, and non-idempotent mutations

openWorldHint

• true for operations that discover or act across a landscape/collection (most tools)
• false for operations on a specific, known resource (e.g., get/create/update/delete a single node by ID)

Quick Reference Matrix

Tool Naming Conventions

• Use snake_case for all tool names
• Start with a verb prefix that matches the operation:

• Keep names concise but descriptive (e.g., list_orgs not list_cloud_foundry_organizations)
• The title field should be the Title Case equivalent of the name (e.g., list_orgs → "List Orgs")

Instructions Field Guidelines

The instructions field tells the AI when and how to use this server.

Simple servers (1-5 tools, single domain)

Use a single sentence:

"MCP server that can be used to send events through the Alert Notification service. Use it to notify your team about operational incidents."

Complex servers (6+ tools, multiple capabilities)

Use markdown with bullet points organized by capability:

"MCP server for managing SAP HANA Cloud database instances through the SAP Service Manager API.\n\nUse this server to:\n- **Instance Lifecycle**: Start, stop, restart instances\n- **Snapshots**: Create, list, delete, and revert to storage snapshots\n- **Configuration**: Update instance size and enable capabilities\n- **Upgrades**: Upgrade instances to newer versions"

Guidelines

• Start with what the server is ("MCP server for...")
• Mention the underlying service or API
• For complex servers, group capabilities with bold section headers
• Mention key parameters users typically need (e.g., "Most commands require an instanceId")
• Use \n for newlines within the JSON string value

Validation Checklist

Before finalizing a generated MCP server definition, verify:

• name matches the intended filename (without .json)
• enabled is set appropriately (true for active, false for draft)
• instructions clearly describes when to use the server
• Every tool has all 10 required fields
• commandId follows the <catalog>-<suffix>:<Name>:<version> format
• name is snake_case and starts with an appropriate verb
• title is Title Case and matches the tool name semantically
• inputReferences uses the correct credential type for the domain (or [] if embedded)
• tags is {} (always empty)
• Hint booleans follow the semantic contract:
  • readOnlyHint: true → destructiveHint: false
  • destructiveHint: true → readOnlyHint: false
  • idempotentHint: true only on safe-to-repeat write operations
• No duplicate tool names within the server
• The output is valid JSON

Reference Examples

Minimal single-tool server (read-only)

{
  "name": "ans-event-producer",
  "enabled": true,
  "instructions": "Use this server to send Alert Notification Service events. Call send_event when you need to publish an alert.",
  "mcpTools": [
    {
      "commandId": "ans-sapcp:SendEvent:1",
      "name": "send_event",
      "enabled": true,
      "inputReferences": ["ans-<<<TENANT_ID>>>:AnsCredentials:1"],
      "tags": {},
      "title": "Send ANS Event",
      "destructiveHint": false,
      "idempotentHint": false,
      "openWorldHint": true,
      "readOnlyHint": false
    }
  ]
}

Multi-tool read-only server

{
  "name": "btp-resource-discovery",
  "enabled": true,
  "instructions": "Use this server to discover BTP resources. All tools are read-only — they list or retrieve data and never modify state.",
  "mcpTools": [
    {
      "commandId": "cf-sapcp:ListCfOrgs:1",
      "name": "list_orgs",
      "enabled": true,
      "inputReferences": ["cf-<<<TENANT_ID>>>:CfCredentials:1"],
      "tags": {},
      "title": "List CF Organizations",
      "destructiveHint": false,
      "idempotentHint": true,
      "openWorldHint": true,
      "readOnlyHint": true
    },
    {
      "commandId": "cf-sapcp:ListCfSpaces:1",
      "name": "list_spaces",
      "enabled": true,
      "inputReferences": ["cf-<<<TENANT_ID>>>:CfCredentials:1"],
      "tags": {},
      "title": "List CF Spaces",
      "destructiveHint": false,
      "idempotentHint": true,
      "openWorldHint": true,
      "readOnlyHint": true
    },
    {
      "commandId": "cf-sapcp:ListCfApps:1",
      "name": "list_apps",
      "enabled": true,
      "inputReferences": ["cf-<<<TENANT_ID>>>:CfCredentials:1"],
      "tags": {},
      "title": "List CF Applications",
      "destructiveHint": false,
      "idempotentHint": true,
      "openWorldHint": true,
      "readOnlyHint": true
    }
  ]
}

Examples

Example 1: Create a simple read-only MCP server

User: "Create an MCP server for listing Cloud Foundry resources"

1. Define server with name, enabled: true, and clear instructions
2. Create tools for list_orgs, list_spaces, list_apps
3. Set readOnlyHint: true, destructiveHint: false for all tools
4. Use cf-sapcp catalog prefix for built-in CF commands
5. Add BtpCredentials input reference for authentication
6. Validate against the checklist

Example 2: Create a full lifecycle MCP server

User: "Build an MCP server for managing HANA Cloud instances"

1. Create tools: list_instances, get_instance, start_instance, stop_instance, restart_instance, delete_instance
2. Apply correct hints: readOnlyHint: true for list/get, idempotentHint: true for start/stop/restart, destructiveHint: true for delete
3. Use hanalm-<<<TENANT_ID>>> catalog prefix
4. Write rich instructions with markdown grouping capabilities by category

Example 3: Create a minimal single-tool event sender

User: "Create a minimal MCP server with one tool to send ANS alerts"

1. Create minimal server with single tool: send_event
2. Set readOnlyHint: false, destructiveHint: false, openWorldHint: true
3. Add AnsCredentials input reference
4. Write simple one-line instructions
5. Use the "Minimal single-tool server" inline example above as a structural guide

Deploying MCP Servers via API

Prerequisites

Set the following required environment variables:

export AUTOPI_HOSTNAME="emea.autopilot.cloud.sap"
export AUTOPI_USERNAME="your-username"
export AUTOPI_PASSWORD="your-password"

For the full list of supported hostnames (emea, aus, apac, amer, ksa), see ../automation-pilot-content-management-via-api/SKILL.md → Prerequisites.

Ensure curl is available in your environment.

The API requires the GenAI permission for all write operations.

List MCP Servers

curl -s -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers"

Get MCP Server by ID

MCP servers have a system-generated id field (e.g. T000414R2-0000001779286543418-1-1) returned in the create/list responses. This is distinct from the human-readable name field. Always use the id for GET, update, and delete operations.

# First, find the ID from the list response
curl -s -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers" | jq '[.[] | {name, id}]'

# Then fetch by ID
MCP_SERVER_ID="T000414R2-0000001779286543418-1-1"
curl -s -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers/$MCP_SERVER_ID"

Create MCP Server

MCP_SERVER_FILE="path/to/my-server.json"
curl -s -X POST \
  -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
  -H "Content-Type: application/json" \
  -d @"$MCP_SERVER_FILE" \
  "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers"

Update MCP Server

Use the system-generated id (not the name) in the URL.

MCP_SERVER_ID="<id-from-list-response>"
MCP_SERVER_FILE="path/to/updated-server.json"

curl -s -X PUT \
  -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
  -H "Content-Type: application/json" \
  -d @"$MCP_SERVER_FILE" \
  "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers/$MCP_SERVER_ID"

Delete MCP Server

MCP_SERVER_ID="<id-from-list-response>"

curl -s -X DELETE \
  -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
  "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers/$MCP_SERVER_ID"

Deploy MCP Server (Upsert Pattern)

Check for existence first — create if new, update if already deployed. Use the system-generated id returned from a previous create or list call:

MCP_SERVER_FILE="my-server.json"
MCP_SERVER_ID="<id-from-prior-create-or-list>"

STATUS=$(curl -s -o /dev/null -w "%{http_code}" \
  -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
  "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers/$MCP_SERVER_ID")

if [[ "$STATUS" == "200" ]]; then
  echo "Updating existing MCP server..."
  curl -s -X PUT \
    -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
    -H "Content-Type: application/json" \
    -d @"$MCP_SERVER_FILE" \
    "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers/$MCP_SERVER_ID"
else
  echo "Creating new MCP server..."
  curl -s -X POST \
    -u "$AUTOPI_USERNAME:$AUTOPI_PASSWORD" \
    -H "Content-Type: application/json" \
    -d @"$MCP_SERVER_FILE" \
    "https://$AUTOPI_HOSTNAME/api/v1/mcp-servers"
fi

Troubleshooting

Error: Duplicate tool names in server Cause: Two tools have the same name field value. Solution: Each tool name must be unique within a server. Use verb prefixes (list_, get_, create_) to differentiate tools that operate on the same resource type.

Error: Conflicting hint booleans Cause: readOnlyHint: true and destructiveHint: true set on the same tool. Solution: These are mutually exclusive — a read-only tool cannot be destructive. Refer to the Quick Reference Matrix to determine the correct combination for the operation type.

Error: Invalid commandId format Cause: Command ID doesn't follow the <catalog>-<suffix>:<Name>:<version> pattern — missing separator, wrong suffix, or missing version. Solution: Verify catalog prefix, suffix (-sapcp for SAP-provided, -<<<TENANT_ID>>> for custom), PascalCase command name, and integer version number.

Error: MCP server tool not activating Cause: Individual tool enabled field set to false. Solution: Set enabled: true on each tool that should be active. Use false only for tools intentionally included but not yet ready for use.