Discover buying intents with Claude Code + TheirStack MCP

Most teams pick their intent keywords by guessing — they sit in a meeting, brainstorm a list of words their software helps with, and ship it. The result is noisy: too generic on one side ("project management"), too narrow on the other, and miles away from the language buyers actually use in their job posts.

The keywords that predict a sale already exist — in the job descriptions of your best customers. This guide shows how to mine them with Claude Code and the TheirStack MCP, check which are already in our Buying Intents catalog, and request the ones that aren't so you can consume them as a clean signal in production.

What you'll get

A short list of high-signal keywords, each one labeled as either:

• Already in the catalog — usable today via the Buying Intent filter and webhooks.
• Missing — send them to us and we add them to the catalog within a few days, after which you consume them through the same API.

Prerequisites

• A TheirStack account with API access.
• Claude Code installed (or any other MCP-compatible client — Claude.ai, Cursor, etc.).
• The TheirStack MCP connected to your client.
• 5–10 ideal-fit accounts (closed-won, late-stage pipeline, or clear lookalikes — not mediocre customers, they pollute the signal).
• Your product / value-prop page URL.

How it works

ideal accounts  ─►  Claude Code (TheirStack MCP)
                       │
                       ├─ reads recent job posts of each account
                       ├─ reads your product page + each account's site
                       ├─ extracts pain-point keywords that overlap
                       └─ checks each keyword against /v1/buying_intents
                       │
                       ▼
            keyword report
              ├─ existing  →  use today via Buying Intent filter
              └─ missing   →  email us to add to the catalog

Step-by-step

claude mcp add theirstack --url https://api.theirstack.com/mcp \
  --header "Authorization: Bearer $THEIRSTACK_API_KEY"

claude mcp list

PRODUCT_URL: https://pivotapp.ai
ACCOUNTS:
  acme.com
  globex.com
  initech.com
  umbrella.com
  hooli.com

You are helping me discover buying-intent keywords for my product.

PRODUCT_URL: {PRODUCT_URL}
ACCOUNTS:
{ACCOUNTS}

Do this in order:

1. Read PRODUCT_URL and summarize, in 5 bullet points, the concrete
   pains my product solves and the manual tasks it automates.

2. For each account, use the TheirStack MCP to fetch up to 20 recent
   job postings (last 12 months). Use the `company_search` tool to
   resolve the domain to a company, then `job_search` filtered by
   that company_id. Read the descriptions.

3. From those descriptions, extract candidate keywords / phrases that
   describe pains, manual processes, or tooling that overlap with the
   pains from step 1. Keep them specific (e.g. "three-way invoice
   match", not "finance"). Discard generic terms.

4. For each candidate, check whether it already exists in the
   TheirStack Buying Intents catalog by calling the MCP tool
   `buying_intent_search` (or the REST endpoint
   `GET /v0/buying_intents?q=<keyword>`). Mark it `existing` (with
   the intent slug) or `missing`.

5. Output a markdown table with columns:
   keyword | status | matched_intent_slug | accounts_seen_in | example_sentence | mapped_to_product_feature | confidence (low/med/high)

Sort by confidence desc, then by accounts_seen_in desc. Skip any
keyword that appeared in only one account with low confidence.

curl -X POST https://api.theirstack.com/v1/companies/search \
  -H "Authorization: Bearer $THEIRSTACK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "buying_intent_filters": { "keyword_or": ["<intent_slug>"] },
    "page": 0,
    "limit": 25
  }'

Why this beats classifying jobs yourself with an LLM

A common alternative is: pull every job that matches a broad filter, then run each description through GPT/Claude to classify it. That works, but:

• You pay LLM tokens on every job — including the 90% that don't match.
• You own the prompt, the eval, and the drift over time.
• The classification is invisible upstream, so you can't use it in saved searches or webhooks.

With the approach above, the classification work happens once, during discovery. After that, TheirStack does the matching on our side — you only pay credits when a job actually surfaces an intent you care about, and you get to use the signal natively in searches, webhooks, and the API.

Tips and pitfalls

• Seed quality is everything. 5 perfect accounts beat 20 average ones. If you don't have 5 closed-won, use 5 accounts you'd bet your quota on.
• Stay specific, not categorical. "procure-to-pay automation" is a signal; "procurement" is a noun. Specific keywords also work better with our per-topic webhook pattern.
• Run it per language / market. Pain-point phrasing doesn't translate 1:1. Do one pass with FR seed accounts, one with EN, one with ES — you'll discover different keywords each time.
• Don't merge categories with topics. A category groups many topics. For webhook monitoring you want one search + one webhook per topic, not per category — see the linked caveat above.

Related guides

• Identifying companies with problems your software solves — the manual version of this exercise.
• Monitor buying intent on a target account list — how to consume the intents once you have them.
• How to set up a webhook
• TheirStack MCP documentation

Case study

• Qonto uses TheirStack to detect companies with high intents — same playbook, applied at scale.