OData MCP Proxy

A config-driven MCP (Model Context Protocol) server that exposes OData and REST APIs as MCP tools. This enables AI assistants such as Claude to query, manage, and monitor SAP backends through natural language.

The server runs on SAP BTP Cloud Foundry and uses BTP Destinations for secure, token-managed connectivity to OData APIs.

Features

• 32 OData entity sets across 6 API categories, automatically registered as MCP tools

• Full CRUD support -- list, get, create, update, and delete operations where the API permits

• OData V2 query capabilities -- $filter, $select, $expand, $orderby, $top, $skip, and $inlinecount

• Navigation property traversal -- dedicated tools for related entities (e.g., iFlow configurations, message attachments, error details)

• Category-based filtering -- enable only the API categories you need via configuration

• Dual transport modes -- Streamable HTTP for BTP deployment, stdio for local Claude Desktop use

• Automatic OAuth token management -- tokens are refreshed transparently via the BTP Destination Service

Related MCP server: SAP ADT MCP Server

Architecture

Claude / AI Assistant
        |
        | MCP Protocol (stdio or HTTP)
        v
 OData MCP Proxy
        |
        | OData V2 + JSON
        v
   OData Client
        |
        | OAuth2 (via BTP Destination Service)
        v
  BTP Destination
        |
        v
 SAP Cloud Integration
   OData Admin APIs

The server resolves a BTP Destination at startup to obtain the Cloud Integration tenant URL and OAuth2 credentials. On each API call, the destination is re-resolved to ensure tokens remain valid. The OData client translates MCP tool invocations into OData V2 HTTP requests and returns structured JSON results to the AI assistant.

Prerequisites

• Node.js 20+ (18+ minimum, 20+ recommended)

• SAP BTP account with a Cloud Foundry environment

• SAP Integration Suite tenant (Cloud Integration capability)

• BTP Destination configured to point to your Cloud Integration tenant's OData API with OAuth2 authentication

• Cloud Foundry CLI (cf) and MBT Build Tool (mbt) for BTP deployment

Quick Start (Local Development)

1. Clone and install

git clone <repository-url>
cd odata-mcp-proxy
npm install

2. Configure environment

cp .env.example .env

Edit .env and set at minimum:

SAP_DESTINATION_NAME=your_ci_destination_name
MCP_TRANSPORT=stdio

Note: For local development with stdio transport, you must have BTP Destination Service credentials available in your environment (e.g., via VCAP_SERVICES or a default-env.json file).

3. Build and run

npm run build
npm run start:stdio

Or use the development watcher:

npm run dev

4. Connect from Claude Desktop

Add the server to your Claude Desktop MCP configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "odata-mcp-proxy": {
      "command": "node",
      "args": ["dist/index.js"],
      "cwd": "/path/to/odata-mcp-proxy",
      "env": {
        "SAP_DESTINATION_NAME": "your_ci_destination_name",
        "MCP_TRANSPORT": "stdio"
      }
    }
  }
}

Using as an npm Package

You can consume odata-mcp-proxy as a dependency in your own project -- similar to how the SAP Application Router works. No TypeScript compilation or build step required.

1. Create your project

mkdir my-mcp-server
cd my-mcp-server
npm init -y
npm install odata-mcp-proxy

2. Add a start script

In your package.json:

{
  "scripts": {
    "start": "odata-mcp-proxy"
  },
  "dependencies": {
    "odata-mcp-proxy": "^1.0.0"
  }
}

3. Add your API config

Create an api-config.json in your project root. The CLI automatically picks it up from the working directory. See the bundled config files for the full format.

{
  "server": {
    "name": "my-mcp-server",
    "version": "1.0.0",
    "description": "My custom MCP server"
  },
  "apis": [
    {
      "name": "my-api",
      "destination": "MY_DESTINATION",
      "pathPrefix": "/api/v1",
      "csrfProtected": true,
      "entitySets": [
        {
          "entitySet": "Products",
          "description": "product entities",
          "category": "master-data",
          "keys": [{ "name": "Id", "type": "string" }],
          "operations": { "list": true, "get": true, "create": false, "update": false, "delete": false }
        }
      ]
    }
  ]
}

You can also use a custom filename with the --config flag:

odata-mcp-proxy --config my-custom-config.json

Or set it via environment variable:

API_CONFIG_FILE=my-custom-config.json npm start

If no config file is found in the working directory, the bundled defaults (SAP Cloud Integration APIs) are used.

4. Configure credentials

For local development, create a .env file or default-env.json with your destination credentials. The env var prefix is derived from the destination field in your config -- uppercase it and replace non-alphanumeric characters with _.

For example, destination "MY_DESTINATION" maps to:

MY_DESTINATION_BASE_URL=https://my-api.example.com
MY_DESTINATION_TOKEN_URL=https://auth.example.com/oauth/token
MY_DESTINATION_CLIENT_ID=...
MY_DESTINATION_CLIENT_SECRET=...

On BTP, use the Destination Service instead (credentials are resolved automatically via VCAP_SERVICES).

5. Project structure

A complete consumer project looks like this:

my-mcp-server/
├── package.json          # start script + dependency
├── api-config.json       # your API configuration
├── default-env.json      # local BTP credentials (gitignored)
├── .env                  # local env overrides (gitignored)
├── mta.yaml              # BTP deployment descriptor
└── xs-security.json      # XSUAA config (if using OAuth)

Deploying to BTP as a consumer project

Since there is no build step, the mta.yaml is straightforward -- just like the SAP Application Router:

_schema-version: "3.1"
ID: my-mcp-server
version: 1.0.0

parameters:
  enable-parallel-deployments: true

modules:
  - name: my-mcp-server
    type: nodejs
    path: .
    parameters:
      memory: 512M
      disk-quota: 1G
      buildpack: nodejs_buildpack
      health-check-type: http
      health-check-http-endpoint: /health
      command: npm start
    build-parameters:
      builder: npm
      ignore:
        - .git/
        - .env
        - default-env.json
    requires:
      - name: my-destination
      - name: my-connectivity
      - name: my-xsuaa

resources:
  - name: my-destination
    type: org.cloudfoundry.managed-service
    parameters:
      service: destination
      service-plan: lite

  - name: my-connectivity
    type: org.cloudfoundry.managed-service
    parameters:
      service: connectivity
      service-plan: lite

  - name: my-xsuaa
    type: org.cloudfoundry.managed-service
    parameters:
      service: xsuaa
      service-plan: application
      path: xs-security.json

The key difference from a standalone deployment: builder: npm is all you need. MBT runs npm install --production, which installs the pre-built odata-mcp-proxy package from the registry. No TypeScript, no custom build commands.

Deploy with:

mbt build && cf deploy mta_archives/my-mcp-server_1.0.0.mtar

Interactive UI Views (mcp-ui)

Beyond plain data tools, the config file can declare interactive UI views: read-only MCP tools that fetch data through the shared OData clients and return a self-contained HTML page as an mcp-ui embedded resource (with the MCP Apps adapter enabled, so the same widget works on MCP Apps hosts like Claude and on classic mcp-ui hosts).

Add a top-level ui array to your API config:

{
  "server": { "name": "my-mcp-server", "version": "1.0.0", "description": "..." },
  "apis": [ ... ],
  "ui": [
    {
      "tool": "UI_SubaccountsOverview",
      "description": "Interactive overview of all subaccounts",
      "uri": "ui://my-server/subaccounts-overview",
      "template": "ui/subaccounts-overview.html",
      "inputs": {
        "subaccountGUID": { "type": "string", "required": true, "description": "GUID of the subaccount" }
      },
      "data": {
        "subaccounts": { "api": "cis-accounts", "path": "subaccounts" },
        "assignments": { "api": "cis-entitlements", "path": "assignments?subaccountGUID={subaccountGUID}", "optional": true }
      },
      "partials": {
        "/*__SHARED_CSS__*/": "ui/_shared.css",
        "/*__SHARED_JS__*/": "ui/_shared.js"
      },
      "frameSize": ["100%", "760px"]
    }
  ]
}

Per entry:

Path placeholders

{param} expands to a validated tool argument. {$...} expands to a fixed, closed vocabulary of derived values — enough for reporting windows and paging without a templating language (there is no eval and no user-defined function):

FMT is yyyymm (202608), date (2026-08-11), or iso (default). N is an integer, the name of a tool input, or that name with one integer offset (months-1) — the offset form exists so an inclusive window ("the last 6 months, including this one") is expressible, and it is the only arithmetic supported.

"usage": { "api": "uas", "path": "monthlyUsage?fromDate={$monthsAgo(months-1):yyyymm}&toDate={$now:yyyymm}" }

Because dates resolve at call time, a view using them is not a pure function of its arguments — expected for reporting windows, worth knowing when caching.

Pagination

paginate repeats the request until the collection is exhausted, a short page arrives, or maxItems is hit:

"users": {
  "api": "xsuaa-scim",
  "path": "Users?startIndex={$offset}&count={$pageSize}",
  "paginate": { "strategy": "offset", "pageSize": 100, "maxItems": 500,
                "itemsPath": "resources", "totalPath": "totalResults" }
}

A paginated source returns a normalized { items, total, truncated, pages } object rather than the raw response — so templates read .items, and truncated tells them the view is showing a capped subset instead of silently under-reporting.

Trimming the payload

The payload is baked into the template and returned as structuredContent, so raw responses reach the model. select keeps only the listed dotted paths of each item, preserving the surrounding envelope:

"subaccounts": { "api": "cis-accounts", "path": "subaccounts",
                 "select": ["guid", "displayName", "region", "state"] }

Templates are full, self-contained HTML/JS pages. The server replaces the token "__DATA__" with the JSON payload:

<script>
  const DATA = "__DATA__"; // becomes { view, params, data: { subaccounts: [...], ... } } — or null in the ui:// template resource
</script>

< is escaped as \u003c in the JSON, so user-controlled strings can never close the script tag. Aggregation and reshaping are the template's job — the server side stays declarative (there is deliberately no templating language or server-side aggregation DSL).

The tool result contains a short text summary (tool name + item counts per data entry), the rendered page as an embedded ui:// resource, and the payload as structuredContent for hosts using render-data delivery.

The UI machinery (and its @mcp-ui/server dependency) is loaded lazily — configs without a ui section skip it entirely.

Progressive Tool Discovery

By default every operation of every entity set becomes its own MCP tool. That is the right thing for a handful of entity sets and the wrong thing at scale: 32 entity sets produce over 100 tools, and the MCP client best practices recommend switching to progressive discovery once tool definitions occupy 1–5% of the context window. Some clients also cap tool counts outright.

Add a top-level discovery block and the entity tools collapse into two stable meta-tools:

{
  "server": { ... },
  "apis": [ ... ],
  "discovery": {
    "mode": "hybrid",
    "alwaysRegister": ["Subaccounts", "cis-entitlements:Assignments"],
    "maxResults": 25,
    "maxFullResults": 5
  }
}

Omit the block and nothing changes — registration behaves exactly as before.

The two tools

search_operations(query, api?, category?, detail?, limit?) — catalog and inspect in one call. detail: "brief" (default) returns name, category, available operations and description; detail: "full" adds keys, navigation/filterable/selectable properties, per-operation method requirements, and concrete path examples.

Two levels rather than the more common three-tool discover → describe → execute split, for two reasons: the spec's own guidance is to "offer multiple detail levels" on the catalog tool, and it saves a round trip when the model already knows what it wants.

An empty or unmatched query returns the whole catalog rather than nothing, with matched: false and a note saying so — a dead end is worse for the model than a list it can narrow. Search is keyword-based with field weighting (exact name ≫ name prefix ≫ category ≫ description), and splits camelCase so sub accounts finds Subaccounts. Embeddings were deliberately not used: it would pull a model dependency into a package that has none.

execute_operation(api, entitySet, operation, path?, navProperty?, body?, headers?) — routes to the same ODataClient, method and path construction as the generated tools, including requiredScope enforcement (the check is shared, not reimplemented).

Because a generic executor has no per-tool schema to reject bad input, it validates routing itself and every failure names the valid options:

execute_operation({ api: "cis-accounts", entitySet: "Subaccounts", operation: "get" })
→ Operation "get" on Subaccounts needs a key expression in "path".
  Keys: subaccountGUID (string). Example path: ('<subaccountGUID>').

Unknown entity sets suggest the API that does have them; unavailable operations list what is available and why; create/update without a body and unknown navigation properties are rejected before the backend is touched.

Schema resources

Discovery also registers one odata://{api}/{entitySet} resource per entity set, returning the same full schema. Hosts that pre-fetch and cache resources can read a schema with no tool round-trip and no context cost until it is read — the 2026-07-28 spec added ttlMs/cacheScope hints to resources/read for exactly this.

Why the tool list never changes

A tempting alternative is registering concrete tools on demand and firing notifications/tools/list_changed. This implementation deliberately does not, for two reasons from the spec: adding or removing tool definitions mid-conversation invalidates the model's prompt cache (the guidance is to "route every call through a single stable meta-tool so the array never changes"), and the 2026-07-28 revision removed protocol sessions so that tools/list no longer varies per-connection. A fixed tool surface is now the conformant design.

Interactive ui views are always registered and never hidden behind discovery — they are few, and they are the entry points the model should prefer.

Programmatic API

The package root exports a start() function, so you can embed the server in your own entry point instead of using the CLI:

// server.mjs
import { start } from 'odata-mcp-proxy';

await start(); // identical to running `odata-mcp-proxy`

To register extra tools or resources on every MCP session, pass registerExtras. It runs inside the per-session factory, after the generated entity tools, API doc resources, and config-driven UI views:

import { start } from 'odata-mcp-proxy';

await start({
  registerExtras(server, ctx) {
    // server: the session's McpServer
    // ctx.clientsByApi: shared ODataClient instances keyed by API name
    // ctx.apiConfig:    the loaded API config file
    // ctx.config:       the environment-derived app config
    server.registerTool('My_CustomTool', { description: '...', inputSchema: {} }, async (args, extra) => {
      const result = await ctx.clientsByApi['my-api'].execute('GET', 'Products', undefined, undefined, extra.authInfo?.token);
      return { content: [{ type: 'text', text: JSON.stringify(result) }] };
    });
  },
});

ODataClient, resolveDestination, createMcpServer, registerAllTools, registerApiDocResources, and the config types are re-exported from the package root as well.

Migration note: if you previously forked the bootstrap (copying the transport/session wiring and deep-importing from odata-mcp-proxy/dist/... to add your own tools), you can delete that entry point: call start({ registerExtras }) for custom tools, and move interactive views into the config's ui section. Deep dist/ imports keep working via the package's exports map, but the root export is the supported surface.

BTP Deployment (Standalone)

When working with the source repository directly (not as an npm dependency), the project includes its own mta.yaml for deployment to SAP BTP Cloud Foundry. The MTA provisions the required service instances (Destination, Connectivity, XSUAA) and deploys the server as a Node.js application using HTTP transport.

npm run build:btp    # Build the MTA archive
npm run deploy:btp   # Deploy to Cloud Foundry

For detailed deployment instructions, destination configuration, and XSUAA setup, see docs/DEPLOYMENT.md.

Configuration

All configuration is managed through environment variables. The server validates configuration at startup using Zod and fails fast on invalid values.

API Categories

Use ENABLED_API_CATEGORIES to restrict which tool groups are registered:

Set to all (the default) to enable every category.

Available Tools

Tools are dynamically generated from entity set definitions. Each entity set produces up to five tools (_list, _get, _create, _update, _delete) plus navigation property tools, depending on what the OData API supports.

Integration Content

Message Processing Logs

Message Stores

Log Files

Security Content

Partner Directory

Tool Naming Convention

Tools follow the pattern {EntitySet}_{operation}:

IntegrationPackages_list
IntegrationPackages_get
IntegrationPackages_create
IntegrationDesigntimeArtifacts_Configurations_list
MessageProcessingLogs_ErrorInformations_list

OData Query Parameters

All _list tools accept standard OData V2 query options:

• $filter -- e.g., "Status eq 'FAILED'"

• $select -- e.g., "Id,Name,Status"

• $expand -- e.g., "Configurations"

• $orderby -- e.g., "Name asc"

• $top -- e.g., 10

• $skip -- e.g., 20

Transport Modes

HTTP (Streamable HTTP)

Used for BTP Cloud Foundry deployment. The server exposes an /mcp endpoint supporting the MCP Streamable HTTP transport with session management, plus a /health endpoint for CF health checks.

MCP_TRANSPORT=http PORT=4004 npm start

stdio

Used for local development and direct integration with Claude Desktop. Communication happens over standard input/output streams.

MCP_TRANSPORT=stdio npm start

Tech Stack

• Runtime: Node.js 20+ with ES Modules

• Language: TypeScript 5.7+

• MCP SDK: @modelcontextprotocol/sdk 1.17+

• SAP Cloud SDK: @sap-cloud-sdk/connectivity and @sap-cloud-sdk/http-client 4.x for destination resolution and HTTP calls

• Validation: Zod for configuration and input validation

• HTTP Framework: Express 4.x (HTTP transport only)

• Logging: Winston

License

MIT