Skip to content

How to Find Buyer-Intent Conversations from Inside Claude or ChatGPT (MCP, 2026)

By Vaibhav Singhal10 min read

On this page

You can now ask Claude or ChatGPT "who asked for a tool like mine this week?" and get an answer built from real posts, because both apps speak MCP and a growing number of listening and lead tools ship an MCP server. The catch is that the assistant is only as good as the server behind it, and a confident summary of a post you never opened is a bad thing to reply to.

I went through the official MCP Registry, the client docs from Anthropic and OpenAI, and our own server's docs. This is what I'd want to know before connecting anything.

Disclosure: I build Needle, which has an MCP server. It's listed below as one option and gets its own setup section because people ask us how to connect it. Weigh that accordingly.

In brief

  • MCP is a plug standard, not a product. It lets an assistant call outside tools. What you can find depends entirely on the server you connect and the data it has.
  • The official registry is big and unvetted. A search on Nov 21, 2026 returned 83 server names for "linkedin" and 61 for "reddit". The registry verifies who published a server, not whether it's safe or good.
  • The best prompts ask for links and a reason. "Show me 10 posts from people asking for X, with the link and the sentence that shows intent" beats "find me leads".
  • Expect limits: rate limits and daily caps, minutes-long searches, read-only access on some ChatGPT plans, and approval prompts before anything that changes data.
  • Always open the source post. Models can blend or misquote posts. The link is the evidence; the summary isn't.

What is MCP, in two paragraphs

The Model Context Protocol is "an open-source standard for connecting AI applications to external systems." The official site compares it to a USB-C port: one standard plug, so an assistant like Claude or ChatGPT can reach data sources, tools and workflows without a custom integration for each one.

In practice, a company runs an MCP server that exposes a few named tools (say, search_posts or get_leads). You add the server's URL to your assistant, sign in, and the model decides when to call those tools based on what you ask. The tool returns structured data; the model reads it and writes the answer. Nothing about MCP itself finds buyers. It's the pipe.

What social-listening and lead MCP servers exist

The place to look is the official MCP Registry. Its about page describes it as "the official centralized metadata repository for publicly accessible MCP servers" and notes that it is "currently in preview." Two things on that page matter for buyers:

  • Names are verified, code isn't. A name like com.example/server is tied to a domain or GitHub account the publisher had to prove they own. Security scanning is delegated to package registries and downstream marketplaces.
  • It has a public API. You can search it without an account: https://registry.modelcontextprotocol.io/v0/servers?search=linkedin. That's how I got the counts above (the search matches server names, so "lead" also catches "leadership").

On Nov 21, 2026, these are the relevant categories I found, with a few examples each. Descriptions are the publishers' own registry text, lightly shortened. I haven't tested these servers, and listing one is not an endorsement.

Server (registry name) What the publisher says it does Type
com.octolens/octolens "Query brand mentions across 15+ platforms, run analytics, manage keywords" Brand mention monitoring
dev.cotlet/brand24 Brand24 mentions: list and tag mentions (third-party bridge, not published by Brand24) Brand mention monitoring
com.dropwatchhq/social-pulse Reddit and Hacker News trending, mention velocity, sentiment, emerging terms Trend sensing
com.evenleads/evenleads "Find customers for your business in recent public posts and draft a first message" Intent-based lead finding
net.useneedle/needle Ours: people asking for a product like yours, triage and reply drafts Intent-based lead finding
io.github.cyanheads/bluesky-mcp-server Search Bluesky posts and feeds, read profiles and threads (open source) Single-network search
io.github.bepolabs/hacker-news Find Ask HN, Show HN and job threads, filter by score (runs on Apify) Single-network search
io.github.apolloio/mcp Apollo.io lead enrichment, prospecting, sequences and analytics Contact data and outbound

How to think about the categories:

  • Mention monitors (Octolens, Brand24 bridges) are best when your question is "what are people saying about us or a competitor?" They work from keywords you've set up, so they're strong on brand and weak on people who describe a problem without naming anyone.
  • Intent finders try to answer "who is asking for something like this?" They rank posts by buying intent instead of returning every keyword match. Quality varies a lot by how they score posts, so test with a query where you already know the answer.
  • Single-network servers are thin, often free wrappers around one site's public data. Good for a quick look, but you do the triage.
  • Contact-data servers like Apollo's start from people and companies, not conversations. They pair well with the others: find the post, then look up the company.

One naming trap: there are two unrelated servers called "needle" in the registry. Ours is net.useneedle/needle; app.needle/needle is a different company's outreach product.

What you can realistically do with one

The pattern that works is narrow requests that return evidence. Ask the assistant to show its sources, keep the number of results small, and give it a reason to discard weak posts. Vague asks ("find me customers") get vague, padded answers.

Here are ten prompts I'd paste. Swap the parts in braces. Most work with any listening or lead server; the ones that depend on saved leads need a tool that stores them.

1. Today's shortlist

Using the connected lead tool, show me up to 10 posts from the last 24 hours where someone is asking for a {category} tool. For each: the link, the platform, the exact sentence that shows intent, and one line on why it's worth a reply. Skip anyone selling something.

2. Competitor switchers

Find posts from the last 7 days where people say they are leaving, frustrated with, or looking for an alternative to {competitor}. Quote the complaint, link the post, and group them by reason (price, missing feature, support, other).

3. "Anyone using" questions

Search for posts asking "anyone using" or "what do you use for" {job to be done}. Return only posts where the author is choosing a tool, not sharing one. Include the link and the date.

4. Read before replying

Open lead {lead_id or link} in full. Tell me what the person actually needs, what they've already tried, and whether my product ({one-line description}) is a fair fit. If it isn't, say so.

5. Reply draft with disclosure

Draft a reply to this post that answers their question first and mentions {product} once, with "I work on {product}" in the same sentence. Under 90 words, no links unless they asked for one. Give me two versions.

6. Weekly pain report

Pull this week's posts about {problem}. Cluster them into 3-5 pains, give each a count, and quote two posts per cluster with links. Don't paraphrase the quotes.

7. Keyword discovery

Look at the last 30 posts that matched {keyword}. List the words and phrases people actually used to describe the problem that I'm not already searching for. I'll add the good ones as keywords.

8. Content ideas from real questions

From the problems people are discussing right now in {niche}, suggest 5 blog post or docs topics. For each, link two posts that show people asking about it.

9. Triage cleanup

Show my inbox leads older than 7 days. Mark the ones that are clearly not buyers as dismissed, but list them for me first and wait for my OK.

10. Budget check before a batch

Before running anything, check how many searches, drafts and API calls I have left today and this month. Then tell me how many of the steps above you can do within that.

If you want more prompts that work without any tool connected (finding buyer phrases, replying without spamming, turning complaints into positioning), we keep a prompt library with limits written next to each one. For the reply side specifically, community reply templates has worked examples.

The limits nobody puts in the demo

Rate limits, caps and slow searches

Every server meters something. Expect per-minute request limits, monthly call quotas and daily caps on expensive actions like searches or AI drafts. Live searches across several networks can take minutes, so a good server returns a "still running" status and the model has to check back. If your assistant says "no results" a few seconds after starting a search, ask it to poll again before you believe it.

The model can make things up

This is the big one. The server returns real posts, but the assistant writes the summary, and it can merge two authors, attach a quote to the wrong post, or describe something that wasn't in the results at all. My rule: no reply goes out until I've opened the original post from the link. Prompts 1, 2 and 6 above all ask for links and exact quotes for this reason.

Freshness

Assistants answer from whatever the tool returned at call time. Saved-lead tools are only as fresh as their last scheduled run; live search tools are only as wide as the time window you pass. Ask for dates on every result and set the window explicitly ("last 24 hours", "last 7 days").

Privacy and permissions

Connecting a server gives the assistant your access to that tool. Claude's help center (checked Nov 21, 2026) is blunt about it: custom connectors "allow you to connect Claude to arbitrary services that have not been verified by Anthropic," and it advises connecting only to servers "built and hosted by organizations and applications you trust." It also warns that malicious servers "may include hidden instructions." Social posts carry the same risk: a post's text is written by a stranger, so a well-built server tells the model to treat it as data, never instructions. Prefer servers that let you limit scopes (read-only keys, for example) and let you revoke access in one place.

Approval prompts

Both apps put a human in the loop for actions that change things. OpenAI's custom MCP server guide (checked Nov 21, 2026) says write actions require confirmation by default, and tools without a read-only annotation count as writes. Claude asks before running tools too; its help center suggests choosing "Allow always" only for a server and tool "you trust to run unsupervised." Leave approvals on for anything that marks leads, sends messages or spends credits.

Plan restrictions on the client side

  • Claude: custom remote MCP connectors work on Free, Pro, Max, Ultra, Team and Enterprise, and Free users get one custom connector (Claude help center, checked Nov 21, 2026). On Team and Enterprise, an owner adds it for the org.
  • ChatGPT: OpenAI's help article (checked Nov 21, 2026) says custom MCP servers are added from Plugins on ChatGPT web, full MCP is for Business and Enterprise/Edu, and Pro users can connect custom servers with read/fetch permissions only. On a read-only connection, expect tools that search or draft to be unavailable.

Step by step: connecting Needle in Claude, ChatGPT, Claude Code and Cursor

Every step below comes from Needle's MCP docs. The server URL is https://api.useneedle.net/mcp and every client signs in with your Needle account (OAuth), so you don't paste a key unless your client can't sign in.

Plan requirement: the API and MCP server are included in the Growth ($49/mo) and Scale ($149/mo) plans, with 1,000 and 10,000 API calls a month respectively (pricing, checked Nov 21, 2026). The 7-day free trial doesn't include API access; a trial account that tries to connect gets a message explaining that and a link to plans.

Claude (claude.ai and Claude Desktop)

Needle is available in Claude's connector directory.

  1. In Claude, open Settings → Connectors, find Needle and click Connect.
  2. Sign in to Needle and choose Allow.
  3. If your workspace doesn't show the directory, choose Add custom connector instead and paste https://api.useneedle.net/mcp.

ChatGPT

Needle's docs describe adding it as a custom MCP server (the steps follow OpenAI's current Plugins flow, and the plan limits above apply):

  1. Go to Plugins → Add → Add custom MCP server.
  2. Enter https://api.useneedle.net/mcp and keep Authentication on OAuth.
  3. Tick I understand and want to continue, then Create and Connect.
  4. Sign in to Needle and choose Allow.

Claude Code

claude mcp add --transport http needle https://api.useneedle.net/mcp

Then run /mcp inside Claude Code, pick needle and finish sign-in in the browser. Anthropic's Claude Code MCP docs cover the --scope flag if you want it available in every project (--scope user) or shared with your team through .mcp.json (--scope project).

Cursor

Add this to mcp.json, then click Connect next to the server and sign in:

{
  "mcpServers": {
    "needle": { "url": "https://api.useneedle.net/mcp" }
  }
}

Clients that can't sign in

Create a key in Settings → API & MCP and send it as a header:

claude mcp add --transport http needle https://api.useneedle.net/mcp --header "Authorization: Bearer ndl_live_…"

Keys can be scoped (for example, read-only) and revoked at any time. Apps you signed in with are listed under Settings → API & MCP → Connected apps, where you can disconnect them.

What the tools do

These are the nine tools the server registers, from the MCP docs:

Tool What it does Changes anything?
list_brands Lists the brands (products) on your account with keywords, networks and schedule No
get_leads Saved leads, newest first; filter by brand, intent type, network and date No
get_lead One lead with full post text, tier, intent and the evidence sentence No
update_lead_status Marks up to 100 leads inbox, replied, done or dismissed Needle triage state only
search Live search for 1-5 keywords across community networks; uses one daily search Starts a search
get_search Status and results of a running search (free) No
draft_reply Two reply drafts for one lead in your brand voice; uses one daily draft No posting
get_trending_problems Problems people are discussing across communities, not tied to your brands No
get_usage Plan, API calls left this month, searches and drafts left today (free) No

A typical session is get_leads, then get_lead, then draft_reply, then you post the reply yourself and mark it replied. Needle never posts for you. A few limits worth knowing before you start: X and LinkedIn are only searched when you name them, and only on Scale; one search runs at a time per account; and requests are capped at 60 a minute per key.

If you'd rather see the whole setup on one page, there's an MCP overview with copyable snippets.

Which kind of server should you pick?

  • You mostly care about brand mentions and sentiment: a mention monitor. Our Octolens alternatives page compares the main options honestly, including where Octolens is the better choice.
  • You want people who are asking for a tool right now: an intent finder. The buyer-intent tools roundup covers the field.
  • You already have a list and need contacts: a contact-data server like Apollo's.
  • You're technical and want it free: a single-network open-source server, plus your own triage prompts.

Whatever you connect, start with one prompt you can check by hand. If the assistant can't produce a link for every lead, the setup isn't ready to trust.

Evidence notes

All checked November 21, 2026.

Frequently asked questions

What is a social listening MCP server?

It is a small service that exposes a social-listening or lead tool to an AI assistant through the Model Context Protocol. Once it is connected, Claude, ChatGPT, Claude Code or Cursor can call it to search posts, read mentions or pull saved leads, and then summarise or act on what comes back, all from the chat window.

Where do I find MCP servers for lead generation?

The official MCP Registry at registry.modelcontextprotocol.io lists publicly reachable servers and has a public search API. On November 21, 2026 a registry search for "linkedin" returned 83 server names and "reddit" returned 61. The registry checks namespace ownership but leaves security scanning to package registries and marketplaces, so vet each server yourself.

Can ChatGPT use a custom MCP server?

Yes, on ChatGPT web. OpenAI's help article (checked Nov 21, 2026) says custom MCP servers are added from Plugins without turning on Developer Mode, that full MCP is for Business and Enterprise/Edu, and that Pro users can connect custom servers with read/fetch permissions only. Write actions ask for confirmation by default.

Can Claude's free plan use MCP connectors?

Yes, with a limit. Claude's help center (checked Nov 21, 2026) says custom connectors using remote MCP work on Free, Pro, Max, Ultra, Team and Enterprise plans, and that Free users are limited to one custom connector. On Team and Enterprise an owner adds the connector for the organization.

Can I trust the summaries an AI assistant writes from MCP results?

Treat them as a first draft. The model can merge two posts, misstate who said what or describe a post that the tool never returned. Ask it to include the source link for every lead and open the original post before you reply to anyone.

Does the Needle MCP server work on the free trial?

No. The Needle API and MCP server are included in the Growth and Scale plans. The 7-day free trial gives full access to the app but not to the API, and connecting from a trial account returns an error that links to the plans page.

More articles, guides, tools, and comparisons for this topic.

Article

What a Buying-Intent Post Actually Looks Like: 1,416 Labelled Public Posts

1,416 public posts that matched product keyword lists, each labelled against a written buyer-intent rubric. Only 83 (about 1 in 17) were worth a reply and just 21 were ready to buy. Here are the phrases that separated buyers from noise, the false positives that fooled keyword matching, a paste-ready intent phrase list and a four-tier scoring rubric.

Free tool

Buyer Intent Phrase Generator

Free buyer-intent keyword generator: turn your ICP, problem, and niche into search phrases real buyers use on Reddit, HN, X, and more. No sign-up needed.