crustdata-tool-selection

The Crustdata MCP server is Code Mode. It exposes list_tools, get_schema and execute, plus feedback (for reporting a problem with Crustdata's own data or tools) when it's listed. Every data operation happens inside an execute script:

// user query: <the user's request, in their words>
const r = await callTool('person_search', { filters, limit: 25 });
// r = { ok: true, data, credits_used? } | { ok: false, status, errorType, message }

• Start every execute script with a // user query: <the user's request> line, or // model query: <what this run does> for your own follow-up step. The server rejects a script without one before it runs, at no charge.
• list_tools, get_schema, execute and feedback (when listed) are MCP tools: call them directly, never through callTool. Inside a script, callTool('list_tools') or callTool('get_schema') is rejected as an unknown tool. callTool only takes the names list_tools returns.
• Write plain JavaScript for execute — any TypeScript syntax (: Type, as, generics) is a parse error that fails the whole run before anything is billed.
• Call get_schema({ names: ["<tool>"] }) before building filters: it returns filterable columns and value formats; add detail: "detailed" for the full column list and a response example.
• in_, nin_, any_of and has_all take an array. Wrap a single value, as in in_(field, ['OpenAI']); a bare string throws.
• A failed callTool returns { ok: false, ... }; it does not throw. Check ok before using data.

Order of preference (cheapest correct path first):

• Company lookup: company_identify (FREE) → company_enrich by crustdata_company_id (2 credits per returned match; enriching by domain/name can fuzzy-match several companies and multiply the charge)
• People search: person_search (indexed DB, ~0.03 credits/result) → only person_search_live (2 credits/result) when the DB returns 0 results or realtime data is required
• Company search: company_search (~0.03 credits/result) → company_search_live (2 credits/result) only for realtime
• Contact info (emails/phones): person_contact_enrich is THE tool — billed per contact type returned, no base fee, up to 25 profile URLs per call; narrow fields (e.g. contact.business_emails) to cap spend. Use person_enrich only for full profile data (work history, education, skills); its preview: true is free
• Job listings: job_search (filters/sorts/aggregations, ~0.03 credits/result) or job_search_live (real-time retrieval by company, 2 credits/result)
• Social posts: social_post_list_live (posts by one person or company) or social_post_search_live (keyword search across professional-network posts only); for X/Twitter, Reddit or any other network use web_search_live
• Web: web_search_live (1 credit/query) and web_enrich_live (fetch a page by URL, 1 credit/page)
• Credits: account_credits (FREE) for the balance; a data-API tool's result carries credits_used for the exact charge of that call (other tools, such as account_credits, return none)
• Discover valid filter values: company_autocomplete, person_autocomplete, job_autocomplete (FREE)
• Bulk lists (thousands of rows): the async batch_* twins (batch_person_enrich, batch_company_search, …) — submit, poll batch_job_get, read via batch_results
• Alerts: watch_create / watch_discovery_create. Send the whole body with preview: true first (free, creates nothing), show the user what it found and what it will cost, then create it in a later execute run. The preview run returns watch_preview_tokens, a list of { tool, dataset, token, reported }; pass the token of the entry whose reported matches the preview you showed as preview_token. The server refuses a create without that token, and refuses it inside the run that minted it. If you change the body after previewing, preview again. watch_preview only shows a sample notification and does not count.

list_tools returns the full annotated tool list with per-tool pricing; trust it over this summary when they disagree.