Amy — a Nostr client that builds itself

Amy is a concept for a Nostr client with an LLM chat (Claude) at its center. The app ships knowing only three things:

1. Nostr — NIP-01 events, relays, and how to read any NIP on demand.
2. NIP-07 signers — it talks to your browser's Nostr extension (Alby, nos2x, …) to sign and publish.
3. How to reach an LLM — it calls Claude (Anthropic), OpenAI, or Gemini (Google) directly from the browser using your own key. Paste any of them; Amy detects which one it is. It can also talk to any OpenAI-compatible endpoint (a local model via Ollama / LM Studio, or a service like OpenRouter, Groq, or Together) when you give it a base URL.

From there, you build the client. Tell the chat what you want to see — "a feed of the latest notes", "a profile card", "a box to publish a note" — and Claude writes a small live view that renders on the canvas. Your project (views, chat, settings) is saved in the browser, so it loads right back up when you return.

There is no server. Everything runs client-side; Nostr relays and your chosen LLM API (Anthropic, OpenAI, or Google) are the only things it talks to.

The UI ships with light and dark themes — use the ☀/☾ button in the top bar to switch. Your choice is remembered (and defaults to your system preference). Themes are driven by CSS variables on :root, which the views Amy builds inherit, so generated interfaces restyle to match automatically.

Run it

It's plain static files — no build step. Serve the folder over HTTP (ES modules don't load from file://):

python3 -m http.server 8000
# then open http://localhost:8000

First-time setup

1. Click Log in (top right). The guided dialog links you to the Anthropic Console (Claude), OpenAI Platform (OpenAI), and Google AI Studio (Gemini) to create an API key. Paste any one — Amy detects the provider from the key (sk-ant-… → Claude, sk-… → OpenAI, AIza…/AQ.… → Gemini) and verifies it live before connecting. The key is stored only in this browser's localStorage and sent directly to that provider's API. (Model and relays live in Settings.) To use a local model or another OpenAI-compatible service instead, add its base URL in the same dialog (e.g. http://localhost:11434/v1 for Ollama, https://openrouter.ai/api/v1); Amy then speaks the OpenAI format to it, and the key may be left blank if the server needs none. The endpoint must allow browser (CORS) access — local servers generally do.
2. (Optional) Click Connect signer to authorize your NIP-07 extension so views can read your pubkey and publish on your behalf. If a Nostr extension is detected, Amy prompts you to do this right after you log in. No extension yet? The same dialog walks you through installing one — it suggests Alby (or nos2x), where you create a fresh Nostr key (or import an existing nsec), then hit Retry. Connecting imports your own relays (NIP-65) so the outbox model routes through the network you actually use — no manual relay setup — and remembers you: on your next visit your profile loads automatically, no reconnect needed. (Use the menu's Disconnect signer to forget it.)
3. Ask the chat to build something.

Why paste a key instead of "Sign in"? Amy has no server to hold a session, so it authenticates with a key you own, kept entirely in your browser. All three providers' inference endpoints allow direct browser access (CORS), so a verified key works immediately.

How it works

File	Responsibility
index.html	App shell: canvas (left) + chat (right) + settings dialog.
js/app.js	Orchestration, project state, the system prompt, and the model's tools.
js/auth.js	Detects the provider from a pasted key (or base URL) and verifies it live.
js/llm.js	Claude, OpenAI, Gemini and OpenAI-compatible API client + the shared tool-use loop.
js/nostr.js	Relay pool (query/subscribe/publish/count/search), NIP-42 relay AUTH, NIP-45 counts, NIP-50 search, NIP-07 signer, NIP-19 bech32.
js/views.js	Runtime that executes a generated view into the page with an api.
js/nsite.js	Export the project as a NIP-5A static website (manifest + icons + OpenGraph), bundling the view runtime so the published site runs.
js/blossom.js	Minimal Blossom client (BUD-02 upload + kind 10063 server list) used by the nsite export.
js/storage.js	Persists the whole project to localStorage.
js/theme.js	Light/dark theme preference (persisted separately, survives reset).

The tools the model has

• read_nip — fetch any NIP's markdown from the nostr-protocol/nips repo.
• query_relays — run a one-shot Nostr query to inspect real data.
• search_relays — run a one-shot NIP-50 full-text search against the user's search relays (kind 10007, falling back to public indexers).
• get_context — read the connected pubkey and configured relays.
• save_view / list_views / delete_view — create and manage the live interfaces shown on the canvas.

The view contract

Each view is JavaScript executed as render(root, api). root is a fresh element to populate; api exposes query, subscribe, publish, count, search (NIP-50 full-text, routed to the user's kind 10007 search relays), signer, nip19, a tiny el() DOM helper, timeAgo(), and per-view getState/ setState. A view may return an unsubscribe function for live feeds, which Amy calls when the view is closed.

Export as nsite (NIP-5A)

The menu's Export as nsite turns your project into a self-contained static website published on Nostr (NIP-5A). A guided dialog lets you pick a named site (a short id under your key) or your root site, edit the title/description, and choose Blossom servers (prefilled from your kind 10063 list, falling back to public defaults). On export Amy:

1. bundles the real view runtime (views.js/nostr.js/…) plus a small bootstrap and your views' code, and generates index.html, a 404.html, a favicon, an Apple touch icon, and an OpenGraph preview image — so the site is complete and actually runs in any nsite-aware client;
2. uploads every file to your Blossom servers (one signer prompt), and
3. publishes a signed manifest (kind 35128 / 15128) to your relays (one more).

You then get the shareable …nsite.lol address. Two caveats: the published site executes your view code on its own domain, and uploads need a Blossom server that allows browser (CORS) access — most public ones do.

Security notes

This is a build-your-own-client concept, so it executes model-generated JavaScript in your page on purpose — that is the feature, not a bug. The code runs with your session's privileges and can sign events through your NIP-07 extension (which still prompts you per signature). Your API key lives in localStorage. Treat this as a personal, single-user tool; don't host a shared instance with these trust assumptions.