Manage your store locator from ChatGPT or Claude: how it works

What MCP is, in plain words

The Model Context Protocol is an open standard that lets an AI assistant use a product's tools directly, with your permission. Without it, an assistant can only tell you what to do: open this page, click this button, paste this address. With it, the assistant does the work itself and reports back.

Handled Locator's MCP server exposes fifteen tools covering everything the locator does: locations, tags, custom buttons, the widget's appearance and wording, and the embed code. Your assistant picks the right tool from what you say, so you never name a tool yourself. You say "which of my locations is closest to Fort Worth?" and the assistant chooses the distance tool, runs it against your real data, and answers.

The connection is scoped to one account: the one you sign in with. The assistant cannot see any other customer's data, and disconnecting removes access without touching anything you built.

What you can say

These are not invented examples. Every prompt below was run, word for word, against a live locator through ChatGPT before this guide was published, and each one worked as a complete instruction with no follow-up needed.

• "What locations do I have in my store locator?" lists them with addresses.
• "Add a new location: Riverside Coffee Uptown, 2800 Routh St, Dallas TX 75201" creates it, geocodes the address, and it is live on the map.
• "Which of my locations is closest to Fort Worth?" ranks by distance, the same way the widget ranks results for a visitor.
• "Add the Flagship tag to my Dallas locations" creates the tag if it is new and applies it in one step. Visitors can then filter by it on the map.
• "Add a Book a table button linking to our booking page on all my Dallas locations, in #C2540E" puts a branded button on every card named, in one instruction.
• "Change the accent colour to a deep green and show me what would change before you apply it" returns a before-and-after list, and nothing changes until you say so.
• "Give me the embed code for my locator" returns the script tag ready to paste into your site.

Connecting from Claude

There is no key to create and nothing to copy besides the endpoint. The whole setup is a sign-in.

1. 1In Claude, open Settings, then Connectors, then Add custom connector.
2. 2Paste the endpoint: https://app.handledlocal.com/api/mcp
3. 3Claude sends you to the Handled Locator sign-in page. Log in with your normal account.
4. 4Approve the connection. You are returned to Claude, and the tools are available in every conversation.

Connecting from ChatGPT

The same sign-in, behind one extra switch.

1. 1In ChatGPT, open Settings, then Apps & Connectors, then Advanced, and turn on Developer mode.
2. 2Add a connector with the endpoint: https://app.handledlocal.com/api/mcp
3. 3Sign in with your Handled Locator account and approve, exactly as in Claude.

What the assistant can do

The fifteen tools group into four jobs. Read tools answer questions; write tools change your locator and are the ones the safety model governs.

The safety model

A locator is a live page on your website, so an assistant is not allowed to restyle it in one move. Changes to how the widget looks and behaves are two-step by design: the assistant first proposes, which returns the exact list of differences and changes nothing, and only applies after you have seen that list and agreed. If someone edited the locator in the dashboard in between, the stale proposal is refused rather than silently overwriting their work.

Every configuration version is kept. A restore does not erase history; it appends a new version, so a rollback can itself be rolled back. Deleting a location is soft: it comes off the map immediately but stays recoverable from the dashboard.

Access has edges too. A connection reaches exactly one account. Key-based connections can be read-only, and a read-only key is refused at the door for anything that writes. Keys are rate limited at 120 reads and 30 writes a minute, so a stuck assistant cannot hammer your locator. Adding locations counts against your plan's allowance exactly as it does in the dashboard, and revoking access takes effect immediately.

What it costs

The MCP server is included on every plan, including the free trial, from day one. That is a deliberate choice: the fastest way to evaluate a locator is to build yours, and describing it to an assistant is the fastest way to build it. When the trial ends you pick a plan to keep the locator serving, and everything you built by chatting is already there.

The REST location API, which developers use for programmatic sync from their own systems, uses the same keys and remains a Pro and Business feature. The distinction is intent: chatting about your own locator is for everyone, machine-to-machine integration is a developer feature.

When chat beats the dashboard, and when it does not

Chat wins wherever the instruction is short and the clicking is long. Bulk work is the clearest case: tagging a dozen locations, or putting the same booking button on every branch, is one sentence in chat and a dozen edit forms in any dashboard. First setup is the second case: describing the locator you want and watching it exist is faster than learning where every setting lives. Restyling is the third, because propose-and-apply shows you the exact consequences of a vague wish like "warmer and rounder" before you commit.

The dashboard still wins for some jobs. Importing hundreds of locations belongs in the CSV importer, which maps your spreadsheet's own columns. Nudging a mis-geocoded pin to the right spot on the map is a dragging job, not a describing job. And reviewing analytics is reading, which a dashboard lays out better than a chat transcript.

The honest summary: the assistant is the fastest editor the locator has, and the dashboard remains the best viewer.