Servicegraph

The branded entry point to ServiceGraph — use whenever the user explicitly names **ServiceGraph** — "use ServiceGraph to…", "what datasets does ServiceGraph have", "search ServiceGraph for…", "look this up in ServiceGraph", "pull contacts from ServiceGraph for these domains", "how many credits do I have on ServiceGraph". ServiceGraph is a multi-dataset platform of metrics-enriched business data for founders — where to launch, who to email, who to hire. This skill explains how to drive the API (api.servicegraph.co / mcp.servicegraph.co) against ANY dataset — discover what datasets exist, discover a dataset's schema and filters, search free brief rows, and unlock contact + metric detail with credits. Dataset-agnostic by design — it discovers everything through the API and never assumes which datasets or fields exist. When the user describes an intent WITHOUT naming ServiceGraph (e.g. "find a PR agency in NY"), defer to the matching specific skill (find-pr-agency, find-marketing-agency, find-law-firm, …); this skill is for explicit ServiceGraph requests and for datasets no specific skill covers yet. Skip non-US firms, consumer/personal services, and individual freelancers.

The generic way to drive ServiceGraph — a platform of metrics-enriched business datasets for founders: where to launch, who to email, who to hire. Use this skill when the user explicitly reaches for ServiceGraph. For intent-first asks that don't name the brand ("find me a CPA firm"), a specific find-* skill is the better fit — defer to it.

There is no single global catalog, and this skill hardcodes nothing about the data. It discovers everything through the API at runtime, so it stays correct as datasets are added, renamed, or re-priced. Discover the datasets from the API, discover each dataset's schema and filters from the API, then search and unlock against it. Never assume a dataset id, a field name, or a price — ask the API.

Two ways to call

Both speak to the same backend; use whichever your harness has.

• MCP server (preferred when loaded) — https://mcp.servicegraph.co, tool names contain servicegraph. OAuth handles credentials in the harness sandbox; no token enters the model context.
• REST — https://api.servicegraph.co, any HTTP client, Bearer-auth with a vk_… key. The universal fallback.

What the API does

Everything except unlocking is free — discover, inspect, validate, and browse as much as you like; only revealing detail costs credits.

Capability	MCP tool	REST	Cost
Find what datasets exist (ids, sizes, prices)	list_datasets	GET /v1/datasets	free
Discover a dataset's schema + filter fields	describe_dataset, list_fields, list_field_values	GET /v1/datasets/:id…	free
Build & validate a filter (or draft one from plain English)	check_filter, translate_intent	…/check, …/translate-intent	free
Search → free brief rows (identity + headline metrics)	search_dataset	…/search	free
Read an already-unlocked row	get_row	GET /v1/datasets/:id/:apex	free
Unlock rows → reveal contacts + full metrics	unlock_rows	POST …/unlocks	spends credits
Check credit balance	get_credit_balance	GET /v1/me/credits	free

The shape is always the same: discover datasets → discover the dataset's schema → search free briefs → unlock the rows the user picks. Rows are keyed by apex domain (stripe.com, not a full URL). Confirm field and value names against the API before trusting a zero-result search — the filter parser accepts invented values silently.

Auth (REST path)

Keys are vk_* tokens the user mints at https://servicegraph.co/profile/api-keys (free credits on signup). The MCP path needs none of this.

Keep the token out of the model context — never read .env/credential files into context, and route authed calls through a shell wrapper so the key flows from the environment into the Authorization header. On 401, ask the user to set SERVICEGRAPH_API_KEY (env or .env.local) and retry; don't accept the key pasted into chat.

Cost & confirmation

Only unlock_rows spends credits, at the per-row price the dataset reports — read it, don't assume it. Unlocks are atomic (a 402 charges nothing) and cached for the dataset's TTL (re-unlocking within it is free). Confirm the cost with the user before unlocking a batch, and check get_credit_balance first if it's large.