Developer quickstart

Zero to first successful Swiggy tool call, self-serve.

This is the self-serve path. You'll ship working agent code that calls a Swiggy MCP tool against a staging endpoint. You will move through account access, framework install, OAuth, and a first tool call.

Building with a coding agent?
These docs are agent-friendly - point Claude Code, Cursor, or Windsurf at them and let your agent read the source of truth. Click here to set it up.

1. Understand what you're building

Every Swiggy MCP tool call goes through the same loop:

  1. Your agent picks a tool from the server's catalogue (e.g. search_restaurants).
  2. The MCP client in your agent framework sends a JSON-RPC call to mcp.swiggy.com/{server}.
  3. The server authenticates the session, runs the tool, returns { success, data }.
  4. Your agent reads the result and decides what to call next.

Three servers, independent per URL: Food (/food), Instamart (/im), Dineout (/dineout).

2. Apply for production access (when you're ready)

You don't need this to start - everything below works on http://localhost without approval. Apply only when you want to take your integration live. Production access is reviewed today, so record a short video of your agent flow working end-to-end and include the link with your application; it dramatically speeds up review.

Apply at /access with:

  • Integration name and organization
  • Redirect URIs for OAuth (exact-match, HTTPS - http://localhost allowed for dev)
  • Which servers you'll call - any of food, instamart, dineout (v1 scopes are mcp:tools, mcp:resources, mcp:prompts)
  • Expected volume and use case
  • A demo video link (Loom, Drive, YouTube unlisted) - or email it to builders@swiggy.in

You'll get staging access; production follows once your staging integration is green. Your MCP client registers itself via Dynamic Client Registration - no client identifier to apply for. See Access for the full flow.

3. Pick a framework

Any MCP-compatible framework works. Pick what you already use; if you're starting fresh, OpenAI Agents SDK and Vercel AI SDK have the lowest-friction MCP story in 2026.

Supported frameworks (recipes in Build an agent):

  • OpenAI Agents SDK (TypeScript + Python)
  • Anthropic SDK - native MCP connector
  • LangGraph / LangChain (via langchain-mcp-adapters)
  • Vercel AI SDK 6 (experimental_createMCPClient)
  • Mastra, PydanticAI, CrewAI, Google ADK
  • Raw MCP client (@modelcontextprotocol/sdk or mcp Python package)

4. Complete OAuth

OAuth 2.1 with PKCE. See Authenticate for the full endpoint walkthrough. Most frameworks handle this for you - you paste the auth URL into your config, a browser window opens for phone + OTP, and you're back.

5. Make your first tool call

Follow the framework recipe at Build an agent. The minimal first call:

const result = await client.callTool({
  name: "get_addresses",
  arguments: {},
});

If you see a user's saved addresses, you're wired up. Now try search_restaurants with an addressId from that response. From there, the world is yours.

6. Build something real

Pick a recipe:

What's next