thumbtack-mcp

Thumbtack access (curl)

Thumbtack's consumer surface is reachable server-side with plain curl — no bot wall, no key, no session. Everything in this skill is anonymous.

Three surfaces, all verified:

1. Search pages — www.thumbtack.com/k//near-me?zip_code=. Results are embedded as JSON in __NEXT_DATA__.
2. Pro profile pages — carry window.__APOLLO_STATE__ (a bare JS assignment, needs brace-matching) plus a clean ld+json business node.
3. GraphQL — app.thumbtack.com/graphql accepts ad-hoc anonymous queries. Introspection is off, so field names come from an Apollo dump.

Setup

None. curl, jq and node (for the bundled extractor) are all you need.

TT=~/.claude/skills/thumbtack/tt-extract.mjs   # adjust to wherever this skill lives
UA='Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/140.0.0.0 Safari/537.36'

Send the browser User-Agent. Requests work without it, but match the browser.

Core calls

Search pros — -L is mandatory (see the resolve rule below):

curl -sSL -A "$UA" "https://www.thumbtack.com/k/house-cleaning/near-me?zip_code=28203" \
  | node "$TT" next \
  | jq '.props.pageProps.frontDoorPage.proListSection.proListResults'

Read a pro profile — the ld+json path is simplest:

curl -sSL -A "$UA" "" | node "$TT" ldjson | jq '.["@graph"][0]'

For credentials, business hours, payment methods and specialties use node "$TT" apollo instead — those live only in the Apollo store.

The extractor takes one mode: next | apollo | ldjson. It reads HTML on stdin, writes JSON on stdout, and exits 3 if that store isn't on the page.

Ready-to-run projections and jq recipes: references/discovery.md. Read it before composing a call — the per-store field paths are all there.

Rules

• Resolve the slug first. Loose service names redirect to a canonical slug (/k/plumbing/… → /k/plumbers/…). Always pass -L, and take the canonical slug from the final URL (-w '%{url_effective}') rather than assuming your guess was right.
• Pick the store by page type. Search pages have __NEXT_DATA__ and no Apollo state; profile pages have Apollo state and ld+json and no __NEXT_DATA__. Using the wrong mode exits 3.
• jq alone cannot read a profile page. __APOLLO_STATE__ is a bare window.x = {…} assignment, not a script tag — that is what the bundled extractor's brace-matcher is for. Don't try to regex it.
• Address the Apollo payload by key prefix, never by literal key: it is servicePage({...the entire input json...}) and changes per page. .ROOT_QUERY | to_entries | map(select(.key|startswith("servicePage")))[0].value
• Never filter JSON-LD on @type == "LocalBusiness". Thumbtack uses the trade-specific schema.org subtype — a plumber's node is "@type":"Plumber", an electrician's is "Electrician". Take @graph[0], or match by shape (a name plus an aggregateRating/address).
• GraphQL answers HTTP 200 with an errors[] body. Check .errors, not the status code.
• servicePage needs its full input object. A partial input returns badRequest even when the query validates. Lift the whole input from the page's own Apollo key.
• Search returns 10 results and no cursor. Deeper paging needs GraphQL.
• Shapes vary by category. priceInfo is a union — a pro may return ServicePagePriceSubsectionNoPrice ("Contact for price") instead of a number. Handle the absent case; don't assume a price exists.
• Read-only. Nothing here writes, books, or contacts a pro.

What this skill does NOT cover

• Your own Thumbtack account (projects, messages, quotes). The login is reCAPTCHA-gated, so there is no server-side password login — that surface needs the browser bridge and is not part of this skill.
• The Partner Platform API (developers.thumbtack.com). Real, documented, and gated behind partner credentials via "Request Access" — a seller surface, not a consumer one.