Teaching Claude to Book a Plumber

A few weeks ago, a Qwikyo engineer typed something into Claude that would have been impossible a year earlier:

"Find me a plumber in Hyderabad for tomorrow morning, budget under ₹800."

Claude searched Qwikyo's task marketplace, surfaced three open tasks matching the criteria, and offered to post a quote — all without leaving the conversation.

This wasn't a demo. It was our MCP integration in production.

What Is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard from Anthropic that defines how AI models connect to external tools and data sources. Think of it as a universal adapter — instead of every AI application building custom integrations with every service it wants to access, MCP provides a shared protocol that both sides agree on.

On one side you have an MCP client — Claude, Cursor, any AI assistant that supports the protocol. On the other side is an MCP server — a service that exposes a set of tools the AI can call.

The AI sees a list of tools with descriptions. It decides which ones to call based on the user's intent. The server executes the calls and returns structured results. The AI folds those results into its response.

That's it. The protocol is deliberately simple. The power comes from what you expose through it.

What We Built

Qwikyo's MCP server lives at mcp.qwikyo.com. It exposes 25+ tools across four verticals:

Task marketplace — search open tasks by keyword and city, get full task details, post new tasks, submit quotes, view your active assignments. Tasks are the core of Qwikyo: someone needs a ceiling fan installed in Banjara Hills, someone else quotes ₹450 to do it.

Marketplace classifieds — search buy/sell/rent listings, get listing details, post your own listings. The OLX-style side of the platform.

Professional services — find tutors, photographers, mechanics, chefs, pandits, and other professionals by skill and location. Check availability and make bookings.

Driver bookings — get fare estimates for point-to-point trips, find nearby drivers, book rides. The Rapido-style vertical.

All of this sits behind OAuth 2.1 with PKCE. Public browsing tools (search, browse, fare estimates) work without auth. Actions — posting a task, submitting a quote, booking a driver — require the user to sign in. Claude prompts for authentication transparently when it's needed.

Why This Matters (Beyond the Demo)

The exciting part isn't the current toolset — it's what the architecture unlocks.

From command line to conversation

Today, most Qwikyo interactions start with a search bar or a category grid. MCP adds a new surface: natural language. A user doesn't need to know how to navigate to "Tasks → Hyderabad → Home Repair → Plumbing." They say what they need, and the agent figures out the rest.

This is especially relevant for India, where users span a huge range of comfort with structured interfaces. Conversational access lowers the barrier significantly.

Agents that work on your behalf

The current integration is human-in-the-loop — you ask Claude, Claude searches, Claude presents options, you decide. But MCP tools are composable. An agent can chain multiple calls: find tasks that match your skills, check which ones you haven't quoted on yet, draft a quote based on your past pricing, submit it — all as one workflow.

We've designed the tools and auth model specifically to make this kind of autonomous use possible when the user explicitly grants it.

The marketplace as infrastructure

Platforms that win the next decade will be the ones that become infrastructure for AI. If Claude can book services on Qwikyo, Qwikyo shows up not just in App Store search results but in AI responses — a fundamentally different distribution channel.

We're early. The MCP ecosystem is still forming. That's exactly when it's worth building.

Technical Decisions

A few things we chose deliberately:

Remote over local. MCP servers can run as local processes on the user's machine or as remote HTTP services. We chose remote. A Cloudflare Worker that's always available, authenticated, and zero-maintenance for the end user is strictly better for a consumer-facing integration.

OAuth over API keys. API keys don't tie to user identity. OAuth does. When an agent posts a task on your behalf, it should be your task — linked to your account, your reputation, your chat history. We use Supabase Auth under the hood, so the same sign-in that works on qwikyo.com works here.

Typed tools with narrow schemas. Each tool has a precise input schema. The AI doesn't get to send freeform JSON — it sends validated, typed arguments. This reduces hallucinated calls and makes error messages actionable.

Public tools without friction. Tools that don't modify data — search, browse, estimate — work without authentication. An AI can pull real task data to answer "what's the going rate for painters in Pune?" without asking anyone to log in.

Connect It

The server is live and works today with Claude, Cursor, and any MCP-compatible client.

Claude Desktop:

{
  "mcpServers": {
    "qwikyo": {
      "url": "https://mcp.qwikyo.com/mcp",
      "transport": "http"
    }
  }
}

Claude.ai web: Settings → Integrations → enter mcp.qwikyo.com

One-line install:

npx @qwikyo/mcp

What's Next

We're actively working on per-user AI agent instances — a dedicated agent per Qwikyo account that learns your preferences, surfaces relevant tasks proactively, and can act autonomously when you give it standing instructions. The MCP server is the first piece of that foundation.

If you're building something on top of this — an agent, a workflow, an integration — reach out. We want to know what people are doing with it.

👉 Explore the MCP server
👉 Post a task on Qwikyo
👉 Browse open tasks

Qwikyo — Get Anything Done, Anywhere in India.