Skip to main content
Deep Research is the heavyweight Source. It runs multiple searches, reads pages deeply, and writes a synthesized answer with citations. Use it for questions that a single search wouldn’t crack. Toggle it from @, send your prompt, and walk away — runs can take a minute or more.
Screenshot coming. Deep Research run with intermediate steps + final synthesis.

​
How it works

Behind the scenes, Foundry runs a fan-out → extract → cite pipeline:
StepWhat happens
SearchThe primary search engine (Exa) returns ~8 high-signal results. If Exa returns nothing, Foundry falls back to Parallel.
ExtractFoundry pulls long-form content from the top pages via Firecrawl.
ReadThe model reads the extracted content into context.
SynthesizeThe model writes a structured answer with citations to every claim.
You’ll see status updates while it runs — “Searching”, “Extracting”, “Reading”, “Writing”. Tab away if you want; browser notifications will ping you when the answer lands.

​
When to use it

  • Market or competitive landscapes — “compare the top 5 prompt-optimization platforms”.
  • Long-form factual digs — “what’s the current state of evidence on cold plunges”.
  • Hard research questions — anything you’d otherwise spend half a day on with tabs.

​
Good for

  • “Summarize every Anthropic paper from 2024 with one-line takeaways.”
  • “Find every YC company in the AI infrastructure space funded in the last 6 months.”

​
Limits

  • It’s slower. Expect a minute or more for non-trivial prompts.
  • It’s more expensive — multiple searches + extractions cost more than a single Web call.
  • The answer is only as good as the sources. Niche topics with thin web coverage produce thinner answers.

​
Tips

  • Be specific. “Compare X, Y, and Z on dimensions A, B, C” outperforms “tell me about X, Y, Z”.
  • Pair with browser notifications so a long run doesn’t tie you down.
  • Open citations to verify any claim that surprises you.

Web

Single-shot search.

Connectors: Firecrawl

The page extractor under the hood.