Expose BoligWatch as an MCP server

[nille]

Motivation

Today Claude Code interacts with BoligWatch by shelling out to python boligwatch.py --peek and parsing the JSON output. This works, but it's a loosely coupled integration — Claude has no structured understanding of what tools are available, what parameters they accept, or what the response schema looks like. It just runs a command and hopes the output is parseable.

Turning BoligWatch into a Model Context Protocol (MCP) server would make it a first-class tool that any MCP client (Claude Code, Claude Desktop, Cursor, VS Code, etc.) can discover and call natively with typed inputs and outputs.

Why MCP over CLI

Concern	CLI (--peek / --json)	MCP server
Tool discovery	Claude must be told what flags exist	Client auto-discovers available tools and their schemas
Parameter validation	Errors surface as stderr text	Invalid params rejected before execution with structured errors
Structured output	JSON blob Claude must parse	Typed response objects the client understands natively
Multi-step workflows	Separate subprocess per action	Persistent server, multiple tool calls in one session
Portability	Only works via shell access	Works in Claude Desktop, IDEs, and any MCP client

Proposed tools

search_listings

Search for rental listings matching criteria. Returns structured listing objects.

Parameters:

• city (string[], optional) — cities to search
• bbox (object, optional) — { min_lat, min_lng, max_lat, max_lng }
• rooms_min / rooms_max (int, optional)
• max_rent (int, optional) — max monthly rent in DKK
• min_size_m2 (int, optional)
• min_rental_period (int, optional)
• pet_friendly / balcony / furnished (bool, optional)
• only_new (bool, default true) — filter to unseen listings only

get_listing

Fetch full details for a single listing by ID.

mark_seen

Mark one or more listing IDs as seen so they don't appear in future search_listings calls with only_new: true.

get_stats

Return current tracker state — how many listings seen, last check timestamp, etc.

Implementation notes

• Use the MCP Python SDK (mcp) for the server
• stdio transport is the simplest starting point and works with all clients
• The existing SearchConfig, Listing, SeenTracker, and fetch_listings() can be reused directly — they're already cleanly separated from the CLI layer
• Keep the CLI entrypoint working alongside the MCP server (they share the same core logic)
• Zero external dependencies is a nice property of the current tool — adding mcp as a dependency is the tradeoff, but it's the only one needed

Example client config

{
  "mcpServers": {
    "boligwatch": {
      "command": "python",
      "args": ["path/to/boligwatch_mcp.py"],
      "env": {}
    }
  }
}

What this unlocks

• Claude can search, filter, and act on listings as native tool calls without prompt engineering around CLI flags
• Works in Claude Desktop and IDE extensions, not just Claude Code terminal sessions
• Composable with other MCP servers (e.g. Playwright MCP for browser automation, Slack MCP for notifications)
• Other MCP clients beyond Claude get access for free