Skip to main content

MCP Setup

@abapify/adt-mcp is a Model Context Protocol server that exposes the same typed contracts as the CLI to any MCP-aware AI client.

  • stdio mode (adt-mcp) is stateless and carries SAP connection parameters per tool call.
  • Streamable HTTP mode (adt-mcp-http) is session-scoped and uses sap_connect / sap_disconnect with Mcp-Session-Id.

See the MCP overview for architecture, or jump straight to the tool reference.

Install​

# Global install (recommended for IDE integration)
npm install -g @abapify/adt-mcp

# Or run on demand with npx / bunx — no install required
npx -y @abapify/adt-mcp
bunx @abapify/adt-mcp

Verify it starts and speaks MCP:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | npx -y @abapify/adt-mcp

You should see a JSON-RPC response listing 100+ tools (discovery, get_object, activate_object, cts_list_transports, …).

Credentials model​

stdio (adt-mcp)​

adt-mcp stdio mode does not persist session state between calls. Each tool invocation accepts connection parameters:

FieldRequiredExample
baseUrl✓https://sap.example.com:44300
username✓DEVELOPER
password✓...
clientoptional100
languageoptionalEN

Most MCP clients let you pin default values via env so the AI doesn't have to repeat them.

Streamable HTTP (adt-mcp-http)​

In HTTP mode, clients should:

  1. initialize and retain Mcp-Session-Id
  2. call sap_connect once (explicit credentials, multi-system config, or local ~/.adt/sessions/<systemId>.json fallback)
  3. run regular tools without re-sending credentials
  4. call sap_disconnect or close the session

See the full deployment guide: Deploying the adt-mcp HTTP server.

warning

Credentials live in plain text inside the MCP client's config file. Use a service account with the minimum authorizations needed, not your personal user.

Claude Code​

claude mcp add adt -- npx -y @abapify/adt-mcp

Or edit ~/.claude/mcp_servers.json manually:

{
  "mcpServers": {
    "adt": {
      "command": "npx",
      "args": ["-y", "@abapify/adt-mcp"],
      "env": {
        "ADT_BASE_URL": "https://sap.example.com:44300",
        "ADT_USERNAME": "DEVELOPER",
        "ADT_PASSWORD": "...",
        "ADT_CLIENT": "100"
      }
    }
  }
}

Restart Claude Code. You should see adt in the MCP panel with all tools listed.

Claude Desktop​

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "adt": {
      "command": "npx",
      "args": ["-y", "@abapify/adt-mcp"]
    }
  }
}

Cursor​

Add to ~/.cursor/mcp.json (or the workspace-level .cursor/mcp.json):

{
  "mcpServers": {
    "adt": {
      "command": "npx",
      "args": ["-y", "@abapify/adt-mcp"],
      "env": {
        "ADT_BASE_URL": "https://sap.example.com:44300",
        "ADT_USERNAME": "DEVELOPER",
        "ADT_PASSWORD": "...",
        "ADT_CLIENT": "100"
      }
    }
  }
}

Reload the Cursor window; the ADT tools appear in the chat tool picker.

VS Code (GitHub Copilot Chat)​

With the MCP-enabled Copilot Chat preview, edit .vscode/mcp.json in the workspace:

{
  "servers": {
    "adt": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@abapify/adt-mcp"]
    }
  }
}

Or add to user settings (settings.json) under "chat.mcp.servers".

Generic JSON-RPC client​

Any client that can spawn a process and exchange JSON-RPC messages on stdio works. Minimal smoke test:

npx -y @abapify/adt-mcp

Then send:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{}}}
{"jsonrpc":"2.0","id":2,"method":"tools/list"}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"discovery","arguments":{"baseUrl":"https://sap.example.com","username":"DEVELOPER","password":"...","client":"100"}}}

Running from source​

Cloning the monorepo and pointing the client at the built binary is useful when tracking main:

{
  "mcpServers": {
    "adt": {
      "command": "node",
      "args": ["/absolute/path/adt-cli/packages/adt-mcp/dist/bin/adt-mcp.js"]
    }
  }
}

Rebuild with bunx nx build adt-mcp after each pull.

First AI prompts to try​

Once the server is connected, ask your assistant things like:

  • "Use the discovery tool against my DEV system and summarize the collections."
  • "Run get_object on ZCL_MY_CLASS and explain the public interface."
  • "List my open transports with cts_list_transports."
  • "Run ATC on package $TMP using atc_run and report only priority 1–2 findings."
  • "Call check_syntax on the current class source and tell me what to fix."

Troubleshooting​

Tool calls return 401 / 403​

The env vars are missing or wrong. MCP clients don't always surface stderr; run the server manually (npx -y @abapify/adt-mcp) to see errors.

spawn npx ENOENT​

The client can't find npx on PATH. Use the absolute path (/usr/local/bin/npx) or install the CLI globally and call adt-mcp directly.

Tools list is empty​

The client connected but hasn't initialized. Restart the client; ensure the server's stdout is not being intercepted by a wrapper script.

Next steps​