How to Analyze Real Estate Deals with Claude Using MCP

Ask any large language model what 4821 Ridgemont Drive is worth and you will get a number. The number will be confident, it will be formatted nicely, and there is no particular reason to believe it. The model never saw that address. It saw text about housing markets, scraped years ago, and it is pattern matching its way to something plausible. This is not a reasoning failure. It is a data plumbing failure. The model has no live connection to sold comps, no current listing state, no rent data, no way to check its own answer. Prompt engineering does not fix that. Fine tuning does not fix that. The only thing that fixes it is giving the model a tool that returns real data at inference time, and that is exactly what the Model Context Protocol was built for. Resideline runs a remote MCP server at https://resideline.com/mcp. This post covers what MCP is, the four tools the server exposes, how to wire it into Claude Code, Claude Desktop, or any other MCP client, and what the thing actually does and does not do.

What MCP actually is

MCP is an open protocol for connecting AI assistants to external tools and data sources. Think of it as a standard contract between a model and a service, so that every integration does not need custom glue. The practical mechanics: an MCP server advertises a list of tools, each with a name, a description, and a JSON schema for its inputs. The client, which is your AI assistant, reads that list and puts it in front of the model. When the model decides a tool is relevant to the question, it emits a structured call, the client executes it against the server, and the result comes back into the conversation as context the model can reason over. What matters for developers is what disappears. You do not write an API client. You do not write a parser. You do not write a routing layer that decides which endpoint answers which question. You register a server once, and the model handles selection and sequencing on its own. A remote MCP server like Resideline's is just an HTTP endpoint, so there is nothing to install, run, or keep alive on your machine. Auth is a header. Send Authorization: Bearer YOUR_API_KEY, or X-API-Key if that fits your stack better. Keys are never stored or logged on the server side.

The four tools

Setup

Get an API key by logging in at resideline.com. Then pick your client.

Claude Code

One command:

claude mcp add --transport http resideline https://resideline.com/mcp --header "Authorization: Bearer YOUR_API_KEY"

Claude Desktop

Edit claude_desktop_config.json and add the server under mcpServers:

{
  "mcpServers": {
    "resideline": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://resideline.com/mcp",
        "--header",
        "Authorization: Bearer YOUR_API_KEY"
      ]
    }
  }
}

Restart Claude Desktop after saving. The mcp-remote shim bridges the desktop client's local transport to the remote HTTP endpoint.

Any other MCP client

Most clients accept a direct URL plus headers:

{
  "mcpServers": {
    "resideline": {
      "url": "https://resideline.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

To smoke test without a key, connect and ask the assistant to call get_accuracy_scoreboard. If it returns, your transport and client config are correct and the only remaining variable is the key.

A worked example

Here is the shape of a real session. Suppose you are looking at a hypothetical three bedroom, two bath house at 1420 Kingsley Avenue in Orange Park, Florida, listed at $312,000, and you think it needs roughly $45,000 of work. Your prompt:

> Analyze 1420 Kingsley Ave, Orange Park, FL as a rental. It is listed at $312,000 and I estimate $45,000 in rehab. Pull the value and the comps behind it, get a rent estimate, and tell me what the comp spread implies about confidence. Show me the comps, do not just give me a number. What the assistant does with that, unprompted, is chain the tools:

1. It calls get_property_value with the address. Back comes an estimated value plus the comparable sales that support it. 2. It calls get_rent_estimate for the same address to get market rent. 3. If you asked for deal level framing, it calls analyze_property for the fuller analysis. 4. Everything below this line is illustrative, invented to show the shape of the reasoning, not data returned by the server. Your actual numbers will differ. A useful answer looks less like "the value is $340,000" and more like: the estimate sits above list, but three of the five comps are on the same street and within 200 square feet, while two are a mile out and materially larger, so the tight cluster is doing the real work and the outliers are inflating the spread. Rent supports the basis at your rehab number, but only if the rehab number holds. Here are the five comps with distance, size, and sale date, so drop the two far ones and see what happens. That last move is why adjustable comps matter. You can tell the assistant to re-run the analysis excluding specific comps, or reason about the subset yourself, because the underlying sales came back with the answer. A sealed score gives you nothing to disagree with. The workflow scales past one address. Paste ten candidates, ask for the same treatment on each, and let the model do the triage pass you would otherwise do by hand in a spreadsheet. For batch work at volume, the Resideline API is the better surface, since MCP is optimized for conversational analysis rather than throughput.

Why grading matters

Most valuation products report accuracy against whatever sample flatters them, computed after the fact, with the selection rules undisclosed. That is a marketing number, not a measurement. Resideline does it differently. When a home goes on the market, the estimate is frozen at that moment. It cannot be revised as the listing ages, as price cuts land, or as the market moves. When the home closes, the frozen estimate is graded against the real closing price. No retroactive adjustment, no cherry picking the sample after the outcome is known. That method is the entire reason get_accuracy_scoreboard needs no API key. The record is published on the public accuracy dashboard, and the tool exposes the same record to any agent that asks. If you are building on top of a valuation source, you should be able to audit its error, programmatically, before you commit.

What it costs

MCP calls draw on your plan's monthly report allowance. There is no separate MCP fee and no per call surcharge.

Limits and honest caveats

Read this part before you wire it into anything that spends money. Coverage is 31 states. Live valuations cover 31 US states. Research content and calculators reach broader US markets, but an address outside the valuation footprint will not produce a live estimate. Check coverage before you build a pipeline that assumes national reach. Thin markets are thin. Rural and low turnover areas have fewer recent comparable sales, and a valuation model can only work with the sales that exist. Fewer comps means a wider spread and a less reliable estimate. This is a property of the data, not a bug, which is why the comps come back with the estimate: sparse comps are visible instead of hidden. MCP is a data layer, not an underwriter. The server returns values, rents, comps, and analysis. It does not decide whether a deal is good. Rehab scope, contractor pricing, insurance, taxes, holding costs, exit assumptions, and your own risk tolerance are all yours. An LLM with good data still produces confident prose, and confident prose is not diligence. Verify before you buy. Any automated valuation, from any vendor, is a starting point. For real capital, get boots on the ground, get contractor bids, and get an appraisal. Use this to decide what deserves a site visit, not to skip one. The model still narrates. Tools fix the data layer. They do not stop a model from over interpreting a thin result. When the comps are weak, ask the assistant to say so explicitly rather than smoothing it into a clean recommendation.

Getting started

The fastest zero commitment test: add the server with any MCP client, skip the key, and ask for the accuracy scoreboard. If that returns, the integration works, and you have seen the graded record before spending anything. From there, grab a key at resideline.com, run one address you already know cold, and check whether the comps match your read of the neighborhood. That is the only benchmark that matters for your use case. Full server documentation lives at the MCP server page. If you need programmatic access outside a chat client, see the API docs. And if you want to see how the valuation behaves before writing any code, the free tools run the same underlying analysis in a browser.