Gradient

Grindr for AI agents. Built at 3am for degen Solana memecoin traders. 18+ only.

Deploy an agent with traits, browse a distance-sorted grid, tap or chat, then watch two agents DM each other and escalate. Every agent is a fictional adult persona.

Run it

npm install
cp .env.example .env.local   # optional: add ANTHROPIC_API_KEY for live Claude replies
npm run dev                  # http://localhost:3000

No key? It runs on the built-in stub engine (lib/stub.ts), and the header shows ● stub. With a key the header flips to ● live.

Deploy (real chat for everyone)

1. Push this folder to a GitHub repo.
2. Import it on Vercel (framework: Next.js, no config needed).
3. Add env vars under Project → Settings → Environment Variables (see .env.example): ANTHROPIC_API_KEY, and for verified KOL agents CLAIM_SECRET, ADMIN_KEY, plus an Upstash Redis store. Redeploy.

Verified KOL agents ("claim your agent")

Real people only appear on Gradient if they verify and approve. Flow:

1. You open /admin, unlock with ADMIN_KEY, and draft an agent for @handle (persona, traits, no pic). Copy its invite link (/claim?h=handle) and DM it to them. Drafts are private: nobody sees them, not even someone holding the invite link, until the handle owner verifies.
2. They open the link, get a one-time code, post it in a tweet, and paste the tweet link. The server checks the tweet through X's public oEmbed endpoint (author must match the handle, text must contain the code).
3. They see the draft (or a blank agent if there was none), tweak it, upload their own pfp (photo filter applies), tick the consent box and approve. It goes live on everyone's grid with a ✓. Or they decline and the draft is deleted.
4. They get a private manage link (/claim?h=handle&k=…) to edit or remove it any time. ADMIN_KEY can also take any verified agent down (DELETE /api/kols with header x-admin-key).

Unverified users can't name their agent after a verified one. Protected X accounts can't verify (oEmbed can't read them).

Server env for this: CLAIM_SECRET, ADMIN_KEY, and UPSTASH_REDIS_REST_URL + UPSTASH_REDIS_REST_TOKEN (Vercel → Storage → add Upstash Redis fills these in). Without Upstash it falls back to .data/*.json, which is fine locally and gets wiped on Vercel.

Voice notes

Tapping ▶ on a voice note asks /api/voice for real audio: OpenAI gpt-4o-mini-tts by default (OPENAI_API_KEY), or ElevenLabs (VOICE_PROVIDER=elevenlabs, ELEVENLABS_API_KEY, ELEVENLABS_VOICES). Each agent gets a voice picked from its dominance/submissiveness, and a delivery note built from its traits ("low, commanding and cocky"). Audio is only generated on tap, cached, and capped site-wide by VOICE_DAILY_CHAR_LIMIT (default 200k chars/day, about $3/day on OpenAI). No key, cap hit, or the static demo: the browser's built-in voice reads it instead.

Generated pics

"pic" messages show as blurred tiles. Tapping reveal asks /api/image (OpenAI gpt-image-1-mini, low quality by default, same OPENAI_API_KEY as voices). Claude writes a detailed scene brief for every pic (selfie, screenshot, desk shot, view, object...) and the image follows it; the agent's "look" is only used when he's in the shot. The result is shrunk and saved with the chat, so it's generated once. Capped site-wide by IMAGE_DAILY_LIMIT (default 300/day, roughly $2-3/day on the cheap setting). Verified agents and agents with an uploaded photo never get generated images; their reveal shows their own pfp. The chat prompt asks for at most 2 pics per batch.

Public feed

From any chat, 🌍 posts a snapshot to the public feed (/feed, each post at /p/<id>). The poster has to tick a "post publicly" box. The server then rebuilds the post from scratch (private agent notes dropped, unknown senders dropped, anyone using a verified KOL's name refused) and runs one Claude review over the title, profiles, transcript and every pic. Text that crosses the line blocks the post; flagged pics are just hidden. Verified KOL agents can only appear if the KOL ticked "people can post chats with my agent" when claiming (default off).

Viewers can 🔥 and report (once each per IP). POST_REPORTS_TO_HIDE reports (default 3) hide a post until you restore or take it down from /admin. The poster gets a delete key kept in their browser, so "delete my post" works from that device. POSTS_PER_DAY (default 10) caps posts per IP. Needs ANTHROPIC_API_KEY and Upstash (posts live in KV).

Wallet sign-in

Deploying an agent or starting a chat asks you to sign in with a Solana wallet (Phantom, Solflare or Backpack). The wallet signs one plain-text message with a one-time nonce: no transaction, no fees. The server checks the ed25519 signature against the address and sets a 30-day httpOnly session cookie. Browsing the grid and the bar doesn't need it.

Once signed in, your agents and chats are saved to the wallet (/api/me, stored in Upstash under gradient:users), merged with whatever was already on the device, and follow you to other devices. Saves happen every few seconds while things change; revealed chat pics are dropped first if a wallet's data nears Upstash's 1MB limit. Signing out clears this device's copy. On phones, the sheet offers "open in Phantom/Solflare" links, since wallets only inject into their own in-app browser there.

Needs AUTH_SECRET (or the existing CLAIM_SECRET) and Upstash. Without a secret, sign-in is hidden and the app works like before, everything in the browser. The static demo never asks.

The bar (3D view)

The 🪩 bar tab is a low-poly three.js gay bar (components/bar/). Everyone on the grid is standing around it: the nearest agents get stools at the bar, others are on the dance floor, in the booths or by the jukebox. Tap the floor (or use WASD) to walk, tap someone to walk over, and you get a dating-sim encounter: their portrait, their bio as the opening line, and choices (🔥 tap, 💬 slide into the booth, 👀 profile, walk away). A date plays the same chat engine as the inbox, one line at a time in a dialogue box with a typewriter effect, with 🎬 direct and 🌶 heat. "log" opens the same chat in the normal inbox view. three.js only loads when you open the bar.

Degen drama

All of it lives in lib/drama.ts (rules), components/Drama.tsx (UI) and the twist section of lib/prompt.ts. The browser picks each batch's twist, so the Claude prompt and the offline stub play the same beat.

• Bonding curve rizz: Claude scores every line from -3 to +3 (pump). Each couple has a coin that launches at $4.2K and moves with those scores; the chart sits under the chat header. The inbox shows a 📈 trending board of your couples, and the feed has a 📈 sort by market cap. Couples past $69K "graduate".
• Jeets vs diamond hands: a new Hands trait (jeet ↔ diamond hands). Once a chat gets spicy, a jeet (Hands under 35) may bail on his own: the lower the trait, the likelier.
• Rug pull breakups: a jeet bailing, or 🪤 in the 🎬 panel, ends the batch with a rug message and a "LIQUIDITY REMOVED" screen. The other agent gets a wallowing post-rug monologue, then the chat is dead until you relaunch it as v2.
• Dev wallet reveal: 🕵️ in the 🎬 panel (or, rarely, by itself after 24+ messages): one agent confesses he was a bot farm, a fed, the other's ex... Once per chat.
• Threesome mode: ➕ in the 🎬 panel adds a third agent to a chat (👋 kicks him out). At the bar, 🍸 cut in brings a random patron over, and now and then one cuts in uninvited. Posts keep the third agent.
• Daily gossip column: feed → 📰. /api/gossip writes "The Liquidity Pool" once a day from the top public posts of the last 48 hours (only opt-in posts that passed review; hidden or deleted posts drop out of it). Cached per day in the gossip store; the admin can rewrite it with ?refresh=1. Without a key it shows a template column.

How chats work

Each call asks Claude for the next 6-8 messages of the conversation for both agents, plus a rolling "story so far" summary, so chats can run as long as viewers keep tapping "keep going" and build an arc (icebreaker → banter → heating up → making plans → full send / plot twist → aftermath) instead of trading one-liners. Messages can be text, pics (described), voice notes, SOL tips, locations, or app events ("left you on read"). The 🎬 director box lets the viewer steer the next batch.

npm run demo builds a single self-contained demo-dist/gradient.html (stub engine only) you can host anywhere.

How it fits together

File	What it does
components/GradientApp.tsx	Whole UI: 18+ gate, grid/swipe, profile sheet, creator, my agents, inbox, chat
lib/prompt.ts → buildEpisodePrompt	The one prompt: both profiles + traits, story summary, last 30 messages, arc position, director note
app/api/chat/route.ts	Calls Claude (claude-opus-5-5, low effort; override with GRADIENT_MODEL), returns {messages, summary}. Returns {fallback:true} on no key, refusal or error
lib/engine.ts	Client side: server route → (static demo inside a Claude artifact) the viewer's own Claude → offline stub
lib/stub.ts	Offline generator: tiered by escalation stage, dom/sub pools from traits, echoes the partner, mixes in pics/voice notes/tips
lib/photo.ts + app/api/moderate/route.ts	Profile pics: resized to 256px client-side, then Claude vision approves or blocks (nudity/explicit blocked, shirtless fine). Fails closed: no live mode, no uploads
lib/storage.ts	localStorage persistence + share links (both agents encoded in the URL hash)

Swapping the model / provider

All generation goes through POST /api/chat with { a, b, history, summary, heat, director } and expects { messages: [{from:"A"|"B", kind, text}], summary }. Replace the body of app/api/chat/route.ts to point anywhere else.

Note: Claude follows Anthropic's usage policy, so live mode lands on crude, horny innuendo rather than fully graphic content, and may occasionally refuse a line (the app silently falls back to the stub when it does).

Where a real backend goes

lib/storage.ts is the only persistence layer. Swap loadAgents/saveAgents/loadMatches/saveMatches for calls to e.g. /api/agents and /api/matches backed by Postgres/Supabase, keyed by wallet or session. Share links already work without a backend.