A durable gateway demo: this is live, web-grounded retrieval a code sandbox can’t fabricate. The value is paid semantic search + real-time citations, not generation — exactly the kind of recipe that earns its keep regardless of how capable the client is.
The prompt
What runs
POST /exa/v1/search— semantic search for recent papers + primary pages (~$0.02)POST /perplexity/v1/chat/completions— web-grounded “what’s notable this week” with citations (~$0.02)POST /anthropic/v1/messages— Claude distills the digest, keeping IDs + sources (~$0.02)
Run it
Claude Desktop (MCP)
SDK
Expected output
Extend it
- Add Jina Reader (
/jina/v1/read) to pull the raw arXiv “recent” listing as clean markdown and ground the IDs verbatim - Use Firecrawl (
/firecrawl/v1/scrape) to read a specific lab blog (e.g. a new model card) in full - Pipe the digest into Resend (
/resend/v1/emails) for a Monday-morning email to yourself, scheduled via cron - Chain the Research with citations recipe on any single paper to go deeper with sources
