Private visual + semantic search
Search everything you own.
Leave nothing behind.
LS Search finds your images, video, audio, PDFs, and notes by meaning — reverse image search, text-to-image, semantic recall — inside a sandbox that holds all data in memory and is destroyed the moment your session ends. No disk, no logs, no remote calls.
Latency and compression figures reflect raw algorithm (Hamming-scan) speed measured on a single core; real end-to-end search speed also includes query embedding and varies with hardware, corpus size, and configuration.
What it does
Reverse image search
Drop an image, get your visually similar ones back — CLIP embeddings, sign-bit quantized for speed.
Text finds pictures
Cross-modal search: type “orange sunset over dunes” and matching photos surface, no tags needed.
Every format
Images, video (keyframe search that seeks to the matching moment), audio (transcript + sound), PDFs, code, notes.
Disposable by design
Your uploads are duplicates held in RAM. Session ends → the whole sandbox and its data are deleted.
Two ways in
People → connect a wallet
One click, no account to create. You get an ephemeral session tenant in its own sandbox. Need durable storage for a large library? Apply for the verified persistent tier.
Agents → x402, pay-per-request
No API key, no signup. Call the endpoint, get an HTTP 402 with payment terms, pay in USDC, retry. Built for the agent economy. See the walkthrough →
Why it's different
- Stateless by default. No persistent index, nothing retained between sessions — a stateless, disposable vector search built to forget. Need durable storage? That's the opt-in persistent tier.
- Nothing persists. Plaintext lives only in memory-backed storage; the sandbox is recycled on exit, idle timeout, or TTL.
- No remote calls from your instance. Models are baked in; the sandbox has no network egress. An outbound attempt is blocked and flagged.
- No logs of your queries. The gateway never records paths, queries, or bodies.
- Self-hostable. Same engine runs in Docker on your laptop or on your own Kubernetes. Self-host guide →
- Energy-efficient by design. Binary Hamming search (XOR + popcount over a 32× smaller index) is memory-bandwidth-bound — it runs on a 1–2 W Raspberry Pi Zero. The retrieval scan costs an estimated ~0.04 J (~12 µWh) per search, on the order of 100,000× less than the LLM call it serves; and because the index is disposable, no server sits burning power on a retained index 24/7.*
*Energy is an estimate for the raw Hamming-scan step on one core; end-to-end search also includes query embedding and varies with hardware and corpus.
Pricing
Human, ephemeral
Wallet sign-in. One disposable sandbox per session. Data held in memory, destroyed on exit.
Agent, per-request
x402 over USDC on Base. $0.03 / ingest. No account. Ephemeral sandbox keyed to the payer.
Persistent
Durable per-tenant encrypted storage for large libraries. Requires identity verification. Apply →