Research that takes more than one search.
An agent run takes one question and does the steps a person would. It searches, checks every row against your criteria, fills the columns you name, and cites what it read. Start one from your code, or ask for one in your assistant.
What a run does.
Criteria on every row
Up to ten conditions, checked account by account. Where a condition maps to a field, the run proves it with a filter instead of judging it.
Columns you name
Up to twenty enrichment columns, filled on every row that matched. A column no dataset can support comes back empty, with a reason.
Accounts you already have
Send up to a thousand accounts to leave out, such as your current customers. Coverage is measured against everything in the index that matched.
Cited insights
Statements about the market or technology behind your question, from research reports, papers and trend records. Each one lists the sources it was drawn from.
A report on the run
A short summary of what came back and the bar that was applied, how much of the market it covered, what was left out and why, and what it cost.
A search or a run?
A run usually costs more than a search and can take minutes, so use it when one search cannot answer the question.
| The question | Use | Why |
|---|---|---|
| “Fintech companies in Berlin” | Search | One question, one corpus, one call. |
| “Fintech companies in Berlin that are hiring sales staff, with their CEO and latest funding round” | Agent run | A condition checked per company, and two columns to fill. |
| “Companies like Adyen and Checkout.com” | Similar companies | You already have the examples. |
Start one.
A run is asynchronous. The API accepts it at once and returns a run id. You then stream its progress or read its results when it finishes. Send an Idempotency-Key with every new run, so a retry does not start a second run or a second charge.
Runs return companies or people. In Claude, Claude Code, Cursor or VS Code you ask in words, and the MCP server starts the run for you.
Set up MCPcurl -X POST https://api.sproutsdata.ai/v1/agents/runs \
-H "X-API-Key: $SPROUTS_API_KEY" \
-H "Idempotency-Key: $(uuidgen)" \
-H "Content-Type: application/json" \
-d '{
"category": "company",
"query": "Fintech companies in Berlin",
"criteria": ["is hiring for sales roles"],
"enrichments": ["CEO name", "latest funding round"],
"effort": "medium"
}'
# 202 Accepted, with a run_id. Stream its progress:
curl https://api.sproutsdata.ai/v1/agents/runs/$RUN_ID/events \
-H "X-API-Key: $SPROUTS_API_KEY"
# Or read the results when it finishes:
curl https://api.sproutsdata.ai/v1/agents/runs/$RUN_ID/results \
-H "X-API-Key: $SPROUTS_API_KEY"You: Find fintech companies in Berlin that are hiring sales staff.
Add a column for the CEO and one for the latest funding round.
The assistant starts a research run with sprouts_research_run
and returns the rows, the two columns, what was left out, and
the sources it cited. Clients that show MCP cards also show the
run's progress, with a Cancel button.Effort and cost.
You choose how deep a run goes. Each effort level has a fixed price in credits, whatever the number of results. The price is held when the run starts and settled when you read the finished run; Automatic holds Maximum's price while it chooses, and is charged the price of the effort it ran. A run that fails, or that you cancel before it finishes, is not charged. Runs made with a test key cost nothing either.
How billing works- minimal
- Up to 20 seconds1.02 credits
- low
- Up to a minute2.125 credits
- medium
- Up to 3 minutes8.5 credits
- high
- Deeper than medium42.5 credits
- xhigh
- Up to 30 minutes85 credits
- auto
- The engine picks, up to highCharged at the effort it selects, never more than Maximum
Try one on your own question.
A free account lets you create API keys and comes with starting credits, and the Research Agent page in the dashboard runs one with no code. No credit card required.