dumpd

a second brain is useless if you have to remember to ask it questions.

💡 inspiration

we hate the copy-paste tax. manually copying text, switching tabs, searching, and organizing files is a brain-rot ritual that kills cognitive momentum.

we wanted something that watches what you're doing and just... hands you what you need. no searching, no switching tabs, no interrupting your flow. the right note, doc, or email shows up on its own because dumpd already knows what you're working on.

this is deeply inspired by poke by interaction co. an AI that lives in the background and helps before you ask. we built that, but for developers.

⚡ what it does

dumpd watches your computer in real-time and surfaces the most relevant things from your personal knowledge base before you even know you need them.

it reads what tabs you have open, what github PRs you're reviewing, what's running on localhost, and even what you're listening to on spotify
🎵 if you put on a specific playlist while deep in a project, dumpd recognizes the song and automatically reloads your workspace and notes from that session
nothing is manual. it all just happens in the background

🛠️ what we built it with

frontend: next.js, typescript, tailwind css, framer motion
backend: hono and bun, supabase postgres
AI search: pgvector, finds the most relevant notes by meaning not just keywords
agentic layer: langgraph.js for orchestrating multi-step retrieval, firecrawl and jina reader for web and document ingestion
redis: keeps everything fast by caching your active context so lookups are instant
claude (anthropic): reads and ranks your notes, understands what you're working on, and decides what's actually relevant to surface
poke (interaction co.): send anything to dumpd through poke via imessage. drop a link, paste a note, forward an email, and it lands directly in your knowledge base and becomes searchable. your messenger is now your capture layer.
custom daemon: a background process we wrote that watches your open windows and browser tabs to keep context fresh

🏆 sponsor tracks

best use of claude: claude is what makes the retrieval smart. it reads your notes, understands your current context, and figures out what's actually worth showing you
best use of redis: redis keeps the context layer fast. without it, every lookup would stall waiting on the database
interaction co. / poke: we built dumpd as a poke-native capture layer. send anything via imessage and it lands in your knowledge graph and joins the ambient retrieval loop. poke handles the input, dumpd handles the memory.

😅 challenges we ran into

cloudflare's website protection kept breaking our browser reading loop, returning empty data instead of page content
our AI tunnel crashed with a 502 error and got stuck in an auth loop for hours
instagram blocked us from pulling video content entirely
spotify's API kept crashing, so we ended up looping mos def instrumentals on repeat for half the hackathon

🎉 accomplishments that we're proud of

got a live browser overlay working that shows retrieved context directly on top of whatever you're looking at
built a search that feels instant and finds the right thing by meaning, not just exact words
the whole thing runs silently in the background. you never have to touch it
shipped a working end-to-end retrieval loop from context detection to knowledge surfacing

📚 what we learned

scraping live websites that use cloudflare is genuinely one of the harder problems in software
building something that runs constantly in the background without breaking requires way more defensive coding than you'd expect
auth flows that span multiple services have a lot of places to go wrong