I Added an MCP Server to MyLeadBots. Here Are the Three Tools, and the Line I Won't Let Them Cross.
I added an MCP server to MyLeadBots so I can run prospect research from Claude, Claude Code, or Cursor without opening the dashboard. The protocol part is straightforward. The part I care about is what the tools are allo
I added an MCP server to MyLeadBots so I can run prospect research from Claude, Claude Code, or Cursor without opening the dashboard. The protocol part is straightforward. The part I care about is what the tools are allowed to claim once they are sitting inside a chat.
I am a lead full-stack engineer in Lahore, and the founder of MyLeadBots. It is a solo-founder product I am building in public. This article is a clear look at the MCP surface that is on the site today: three tools, a bearer token, and a hard line on what a score means.
What MyLeadBots does
MyLeadBots is software for small web agencies and freelance website builders who sell websites and quote-request improvements to local businesses. You pick a city and a niche, or draw an area on the map. The product finds businesses live from Google Maps, checks their public presence, and prepares reviewable proof: an audit report, a concept site, a white-label intake form, and pitch drafts for email and WhatsApp.
It pulls name, phone, website, rating, and review history. Rows are deduplicated, and chains are filtered out. There are no purchased lists. The data comes from Google Maps, Lighthouse, the businesses' own websites, and public social profiles. Nothing is scraped from private sources.
The pitch is supposed to start from a specific observation, not from a generic "we build websites" message. That only works if a person checks the evidence before anything is sent. Putting the same workflow behind a chatbot does not change that.
What MCP is, in brief
The Model Context Protocol is an open-source standard for connecting AI applications to external systems. Official docs describe MCP servers as programs that expose capabilities through a standard interface. The building block I am using here is tools: named functions a model can call, with a defined job, rather than a free-form scrape of my product.
I am not going to walk through implementing a server from scratch. The official getting-started and server-concepts pages already cover that. This post is about the product surface I actually shipped.
The three tools
The public homepage lists three tools.
discover_leads finds businesses live from Google Maps. The example on the site is "Find 20 dentists in Houston." Results are deduplicated, and chains are filtered.
get_leads brings the public details into the chat: names, phones, websites, ratings, and review history.
audit_lead runs the deep audit the product already prepares: an opportunity score, a SWOT, a pain summary, and ready-to-send pitch scripts.
The homepage describes it this way: connect once with an API key, and your chatbot can pull live Google Maps leads and run the full multi-agent audit without opening the dashboard. I am quoting that as product wording. I am not going to unpack a private agent architecture here, and I will not add tool names or parameters that are not on the public page.
The intended loop is short:
flowchart LR
A["discover_leads"] --> B["get_leads"]
B --> C["audit_lead"]
C --> D["A person reviews the evidence"]
D --> E["Pitch only if the findings check out"]
The site also shows an illustrative conversation. In that sample, someone asks for ten med spas in Austin with weak websites, and an audit of the best one. The sample reply flags a missing online booking page, returns an opportunity score, and then says the findings are checks to verify, not measured AI visibility. It asks whether to review the report and draft a pitch. That conversation is labelled illustrative on the page. It is a walkthrough of the intended loop, not a report on a real business or a real result.
How you connect
The endpoint on the site is:
https://myleadbots.com/api/mcp
Authorization: Bearer mlb_••••••••
The product says this works with Claude, Claude Code, Cursor, and any MCP-compatible assistant. You connect with an API key. I am not going to invent a client config file for a specific app, because the public page does not publish one. If you use Claude or Cursor, use that app's own remote-server setup and point it at the URL and bearer token above.
The line I will not let the tools cross
Each audited business is scored 0–100 across its Google Business Profile, website speed and SEO, online booking, review velocity, social activity, and buying signals. The product is explicit about what that number is not: scores summarize signals for review. They do not establish demand, budget, or the chance of winning a client.
A missing page or booking link is a reason to investigate, not proof that an owner wants to buy. A concept-site preview does not measure conversion or promise search rankings. The sample pitch opener on the site asks a question instead of making an accusation: "I could not find a booking link on the page I checked. Is there a separate page patients use?"
That disclaimer has to survive the move into a chat window. A model that can call audit_lead can also overstate the output: treat a score as a win probability, treat a missing booking link as a ready-to-close deal, or treat a generated concept site as a live result. The MCP server does not change the underlying rule. The artifacts are there to be reviewed. The message should not go out until a person has checked the business and the findings.
Discovery works in any city, because it runs on Google Maps. The product and the pitch scripts are built for English-speaking markets first: the US, the UK, Canada, and Australia. That limit belongs in this article too. A chatbot that can search a city is not the same thing as scripts that are ready for every language and market.
I also will not attach customer results to this write-up. I have not published reply rates, client counts, or testimonials, and I will not invent any here.
What I would tell you to take from this
If you are adding an MCP server to a product, a small tool list is a feature. These three tools map to the workflow I already ask people to run by hand: find businesses, inspect the public details, then audit and prepare proof. The protocol makes that reachable from a chatbot you already use. It does not make the output safe to send unread.
The useful design question is not "can the assistant do this without opening the dashboard?" It is "what should the assistant be forbidden from implying once the tools have run?" For MyLeadBots, the answer is already on the site. A score is a reason to look closer. It is not a prediction.
If you want to try the product itself, it is at myleadbots.com. New accounts get 20 free credits, and you do not need a card.
If you were adding an MCP server to a tool that prepares outreach, what is the one claim you would refuse to let the tools make?
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.