Intent

Intent is an AI helper for people using AI for the first time, especially older adults. It spots scams, asks before spending, and keeps its memory and rules on your side. Live: https://intent.microvalley.io (no sign-up). THE PROBLEM Older adults are the main target of "grandson in trouble, buy gift cards now" scams, and they are also the people chat-style AI serves worst: small text, long replies, no clear next step, and no guardrails on actions that cost money. WHAT IT DOES - One sentence to start: type or tap-and-speak; results can be read aloud. - Buttons, not prompts: each request becomes its own task with a status on the Today screen. The model's moves are rendered as cards: a question with big answer buttons, a plan to accept or change, a result, or a purchase that needs approval. - Scam check: paste a suspicious message and get a plain verdict, the warning signs, and what to do next (e.g. call your grandson on the number you already have). - Guardrails enforced in server code, not the prompt: gift cards, prepaid cards, money transfers, loans and investments are on a never-do list; per-purchase and monthly limits are re-checked when the user taps approve. Even if a scam message manipulates the model, the server refuses. - Memory you can see and delete; card numbers, passwords and codes are never stored. Three text sizes, Chinese and English, large touch targets. A TINY, SELF-CONTAINED AGENT RUNTIME The whole app is one executable (built with Bun): web UI, API, the Muse Spark agent loop, a TinyBase in-memory database persisted row by row to built-in SQLite, live updates over server-sent events, anonymous sessions plus optional username/password accounts (scrypt), and a PocketBase-style /_/ admin page (first-run setup link, health, users, request log, one-click backup). No Docker, no external database, no hosted backend. The live demo is that single file on a free-tier Google Cloud e2-micro (1 GB RAM) behind a Cloudflare tunnel, using about 55-80 MB of memory. Builds for Linux x64/arm64, macOS and Windows are published (Linux x64 tested). VERIFIED On the live site with the real model, a fictional gift-card scam text gets a clear "this is a scam, don't buy" verdict with reasons and next steps, and a birthday-message request returns a finished result. 30 desktop/mobile browser tests and a 24-check end-to-end test of the binary pass in CI (these use a scripted test model and test the flow, not model quality). BUILT DURING THE HACKATHON An earlier UI and task-flow prototype existed before the event. During the hackathon I added the Muse Spark integration, the server-side purchase policy and approval enforcement, sessions and accounts, cancellation and restart recovery, the single-file runtime (TinyBase + SQLite, admin page, live updates, backups, cross-platform builds), the free-tier deployment and the tests. LIMITS Intent does not pay, order, send messages or run background reminders. The model can miss some scams. Demo data is fictional. Next: a local Llama option, a documented external API and an Agent-to-Agent endpoint, and usability sessions with older adults.