Get the input and confirm the column. Accept a CSV/TSV, plain-text list, or names pasted in chat. For a CSV, name the company column explicitly. Over ~40 names, tell the user each name is one HTTP call and confirm before starting.
Resolve via the free Clearbit autocomplete endpoint — https://autocomplete.clearbit.com/v1/companies/suggest?query=<name>. No key, no auth, up to 5 matches best-guess first, [] on no match. It is an undocumented typeahead, not a contracted API: keep volume modest, pace requests, and read references/endpoint-behavior.md before your first batch — it covers the two failure modes that silently corrupt lists.
For one or two names, query directly with your agent's web-fetch tool and report matches inline, flagging when several results share a name.
For a batch, use the resolver script in references/resolver-script.md. It runs plan → fetch → merge: dedupes on a normalized key, strips legal suffixes before querying, then scores every row. If the environment blocks direct HTTP, the script's plan output lists the URLs so you can fetch them with your agent's web-fetch tool (5–8 in parallel per batch) and feed the responses back to the merge step.

Score every row. Confidence labels and what they mean:

confidence	meaning	review
`exact`	name and domain both match, ranked first	no
`strong`	exactly one result's domain matches the name	no
`plausible`	same slug on several TLDs; picked the gTLD	spot-check
`ambiguous`	several same-named companies, different domains	yes
`weak`	nothing lined up; top result is a guess	yes
`none`	empty response — not in the index	yes

Report, don't just deliver a file. Give counts per confidence level and list the specific rows needing review with their alternatives. If weak plus none exceed a third of the list, say so plainly — the input is likely full of internal shorthand, DBA names, or companies the index doesn't cover. For those rows, offer (don't auto-run) a web search on "<company name>" official site as a second pass.
Hand off. Domains are the join key for enrichment and CRM dedupe. For people-level data (titles, emails, profiles), pass the resolved domains to a real enrichment step — this endpoint returns name and domain only.

What good looks like

The best operator distrusts the top result by default. This endpoint always returns something for common words — "Clay" returns five Clayton-somethings and no Clay — so a bare domain list with no confidence column is malpractice, not a deliverable.
The common mistake is querying names as-is. Legal suffixes wreck results: "Stripe Inc." returns stripe-inc.jp (wrong company), "Bright Data Ltd" returns nothing, while "Stripe" and "Bright Data" both resolve cleanly. Strip trailing Inc/Ltd/LLC/GmbH/Corp/Holdings/Group before every query.
Good output is auditable: every original column preserved, every row labeled, alternatives shown for ambiguous rows, and the review burden quantified up front ("41 exact/strong, 6 need review, here they are") — never a 30% weak-match rate buried in a file.

Rules

MUST attach a confidence label to every resolved row — never hand back a bare name→domain list.
MUST record empty responses as explicit no-matches, distinct from rows that were never fetched — those are different problems with different fixes.
NEVER write weak, ambiguous, or none rows into a CRM unattended; resolve them with the user first.
NEVER present this endpoint's output as authoritative company data or promise logos from it — it is a best-effort typeahead index.