Search, read, synthesize — in one pass
ResearchAI searches the live web, reads what it finds, and writes up a structured report with sources attached. No tabs, no skimming.
How it works
Four steps, same order every time.Top 5 results, 10s timeout.
Your topic goes through Jina's search API, which returns the most relevant pages on the live web right now.
Run in parallel, 8s each. Failures are reported, not hidden.
Each result is fetched and parsed into clean text by Jina's reader — no nav bars, no ads, no boilerplate.
Page text is fenced as untrusted data.
Groq (gpt-oss-120b) reads every page that loaded and drafts a summary plus the findings worth keeping.
Markdown on the clipboard.
You get a structured report: summary, key findings, and the exact sources it read — one click to copy.
hover a card to read what each stage actually does
Why
A summarizer you can audit.Shows its work
Every source is listed with whether it actually loaded. Pages that failed are shown as unreachable, not quietly counted.
Built for speed
Groq's inference means the whole pipeline, search through synthesis, usually finishes in well under a minute.
Copy-paste ready
One click turns the report into clean Markdown, ready to drop straight into your notes.
Don't know where to start?
Run one of these, or type your own.FAQ
What it does, and what it doesn't.No, and it doesn't claim to be. It's a fixed three-stage pipeline: search, read the results, synthesize once. There's no planning loop and no tool selection. It's called an agent in a lot of marketing; here it's a script that runs the same way every time — which is the reason it's predictable.
Yes. The route streams each stage over Server-Sent Events as it happens, including whether each individual page loaded. Nothing in the progress trace is on a timer — if a stage is slow, you watch it be slow.
It's listed in the report and marked unreachable, and its content is excluded from synthesis. If every source fails, you get an error rather than a report written from nothing.
Groq (gpt-oss-120b), running on Groq's inference API. Page text is fenced inside untrusted-data tags before it reaches the model, which blunts the obvious prompt-injection path but doesn't eliminate it — treat findings from an unfamiliar source with the same suspicion you'd give the source itself.
Trust it about as far as you trust the five pages it read, which are all listed with links. It's a reading shortcut, not a fact-checker. Click through before you cite anything.
Start reading one report.Open the agent