Sprouts Data MCP

Forty-three million companies, two hundred million people and seven document collections behind one URL - and a citation on every field, including the fields it does not have.

Sprouts Data publishes its platform corpus over MCP at https://dashboard.sproutsdata.ai/mcp. A bare initialize returns HTTP 401 with the JSON-RPC error -32000 reading No credential. Send your Sprouts API key as an X-API-Key header (or Authorization: Bearer ...). Keys are issued in the dashboard; sk_dsp_test_ keys hit the same corpus at zero credit cost. That last clause is the reason this entry is catalogued with evidence rather than a card: a test key reads the real corpus, so the tool list can be driven before anyone pays.

Server type: Remote (Streamable HTTP)
Auth: sk_dsp_ API key via X-API-Key or Authorization, or OAuth 2.1
Endpoint: https://dashboard.sproutsdata.ai/mcp
Surface: 9 tools, 6 prompts, 8 resources, inline cards (vendor counts)
Corpus: 43M+ companies, 200M+ people, 470M affiliation edges
Scopes: dsp.search.read, dsp.entities.read, dsp.agents.write, dsp.usage.read

Why This Matters for Operators

Most business-data connectors answer one question: who do I sell to. They return rows and a confidence score, and the operator has no way to tell a stale record from a live one, or a missing company from a company the vendor never indexed.

The root cause is that a record set without provenance is only as trustworthy as the vendor's silence about its gaps. This corpus is shaped differently. Records carry per-field provenance, and the vendor documents "honest misses" as a product behaviour rather than an error state: sprouts_entity_lookup charges nothing when a record is not found, and sprouts_coverage returns a not_in_index[] array naming the gaps in plain language. The missing data is published deliberately, which is what makes the present data worth acting on.

Search also runs across nine categories, not one: companies, people, patents, news articles, trends, market reports, funding rounds, research papers and job openings. That changes the questions a chat can answer. Not only which companies fit a profile, but what a sector has been patenting, who raised in it, what the published research says, and whether the hiring pattern moved before the funding did.

Tools and Capabilities

Twelve tools are documented on the vendor's tools page. The server landing page still says nine, so the twelve below is the reading of the page that lists them individually.

Tool	What it answers	Credits
sprouts_search	Describe who to find, across all nine categories	0.5 or 0.595 first 10 results, then 0.08 or 0.085 each
sprouts_similar	Name example companies, get ranked look-alikes	2 first 10, then 1 per next 10
sprouts_entity_lookup	Enrich identifiers you already hold, in batch	1 per record found, misses free
sprouts_relationships	Two-sided questions over the affiliation graph	2 first 10, then 1 per next 10
sprouts_org_chart	The shape of one company's team, as counts	1 per call, free when nobody matches
sprouts_research_run	One search cannot answer it	1.02 to 85 by effort, charged only on success
sprouts_lists / sprouts_list_edit	Saved lists, CSV export, add and remove rows	Free
sprouts_coverage	What the index holds, and what it does not	Free
sprouts_usage	What this has cost, and replaying prior results	Free
sprouts_feedback	A result was wrong, missing or stale	Free
sprouts_top_up	Opens a payment page, never charges in chat	Free

org_chart returns counts and never names, which is the right default for a headcount question and the wrong tool for a person question. list_edit is the only destructive tool in the set and is annotated as such, so a client that surfaces destructive hints will warn before a saved list is changed.

What a Query Costs

Metering is in credits and the prices are published per tool, so a run can be budgeted before it happens. The two shapes worth knowing: sprouts_search is cheap at the first page and steep on the tail if a client pages deep, and sprouts_research_run is the only tool whose cost is open-ended by design, priced by effort and charged only when the run succeeds. An asynchronous run is submitted with sprouts_research_run and resumed by run_id, so a long research brief does not hold a conversation open.

Connecting

1. Create an sk_dsp_ key in the dashboard, or use an sk_dsp_test_ key to read the same corpus at zero credit cost.
2. Add https://dashboard.sproutsdata.ai/mcp as a custom connector.
3. Send the key as X-API-Key or Authorization: Bearer, or sign in over OAuth 2.1 if the client can only take a URL.
4. Ask a coverage question first: sprouts_coverage is free and tells you what the index holds before you spend a credit.

In ChatGPT Business or Enterprise and Claude Team or Enterprise, a workspace administrator may need to enable custom connectors before any individual can connect an account.

Limits Worth Knowing

Keys are per deployment and the failure is indistinguishable. Sprouts runs more than one deployment, each with its own dashboard, data and billing, and a key from the wrong one fails as 401 auth.invalid_key - the same error as a revoked key, with nothing in it naming the environment. When a working config stops working, check which deployment the endpoint belongs to before anything else.

Two versioning facts are worth pinning: the data contract is v1 and changes inside it are additive, and the tool names are frozen (sprouts_search, sprouts_similar, sprouts_entity_lookup, sprouts_relationships, sprouts_research_run, sprouts_coverage, sprouts_usage, sprouts_feedback, sprouts_top_up are all named as fixed). The protocol revision is negotiated per session, so there is nothing for a client to pin.

This pass verified the authorization boundary, the credential contract and the published tool and pricing pages. It did not drive tools/list, because that requires a key. The nine-tool figure on the landing page against twelve on the tools page is recorded rather than resolved.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

See Also

• External MCP Server Catalog
• All sweep reports

FAQ

What does the Sprouts Data MCP server do?

It exposes a business corpus as tools inside Claude, ChatGPT, Cursor or any MCP client: 43M+ companies, 200M+ people and seven document collections covering patents, news, trends, market reports, funding rounds, research papers and job openings, with searching, look-alike accounts, batch enrichment, a relationship graph, saved lists and an asynchronous deep-research run.

Does it need an API key?

Yes. It accepts a Sprouts sk_dsp_ key as an X-API-Key header or as a Bearer token, and also supports OAuth 2.1 for clients that can only take a URL. Test keys prefixed sk_dsp_test_ read the same corpus at zero credit cost.

How is usage billed?

In credits, priced per tool and published per page of results. Searches cost 0.5 or 0.595 credits for the first ten results and 0.08 or 0.085 for each additional result. The asynchronous research run costs between 1.02 and 85 credits depending on effort, and is only charged when the run succeeds.

Can the assistant see data the vendor does not have?

It says so explicitly. sprouts_coverage reports what each corpus holds plus a not_in_index[] list of the gaps, and a lookup miss is charged at zero, which means the absence of a record is a first-class answer rather than a silent one.

What happens if a key stops working?

The most common cause is the wrong deployment. Each Sprouts deployment has its own dashboard, data and billing, and a key from another one fails as 401 auth.invalid_key, which reads exactly like a revoked key and names no environment.