GameGuide

The spoiler-safe game guide you can talk to.

Ask about any game: a boss, a build, a quest, a sale. GameGuide looks it up live, lays the answer out like a strategy-guide page with its sources, and keeps every story beat past your progress behind a bar until you choose to look.

▶ Try it — no account needed · Add to Discord · Top.gg · Security

---

Why it exists

A general chatbot will answer a game question, and it will happily spoil the ending while it's at it, quote last year's patch, or make up an item location. GameGuide is built around the three things a player actually needs from a guide:

• Help from where you are, and nothing past it. Tell it "I just beat Margit" once. Story answers stop at that line; anything beyond it is hidden until you click.
• Current information. Every answer is researched when you ask, from game wikis, official patch notes and web search, with the sources listed so you can check.
• An answer you can act on. Laid out like a guide page (steps, items, numbers), not a wall of chat.

What you can do

Feature	What it does
🛡️ Spoiler Shield	/progress Elden Ring : beat Margit, or just say it. On by default; /spoilers off once you've finished.
🧭 Missables	/missables lists what you can still lose for good (items, questlines, trophies) from where you are, in the order you'll reach them.
🔎 Live research	Wikis, official patch notes and web search, read at the moment you ask, then cross-checked against each other.
🖼️ Screenshots	Attach up to three per message: a build screen, a map, a puzzle.
💸 Deals	/price gta 5 shows the best current PC price across stores and how it compares with the all-time low. It understands nicknames like bg3.
🔗 Spoiler-safe sharing	Share any answer as a link. Spoilers stay hidden for your friend, and the answer travels inside the link, so nothing is stored.
🎨 The page becomes the game	Name a game and the chat takes on its Steam art and a colour sampled from it.
🕶️ Stealth	/stealth starts a conversation that is never saved, remembered or traced.
📡 Discord + Watchtower	The same guide in your server via /ask or an @mention, plus /watch add to post a game's patch notes (summarised) and big deals into a channel.
📱 Installable	Add it to your phone's home screen and keep it open as a second screen while you play.

Free on the web, no account needed. Signing in with Google keeps your history.

How an answer is made

flowchart LR
    A[Your question<br/>+ screenshots] --> B[Understand<br/>which game, what kind<br/>of question, how recent]
    B --> C[Research<br/>wikis · patch notes ·<br/>web search · price data]
    C --> D[Cross-check<br/>sources against<br/>each other]
    D --> E[Answer<br/>from the best available<br/>model right now]
    E --> F[Spoiler Shield<br/>hide anything past<br/>your progress]
    F --> G[Guide-style answer<br/>with sources]

The whole pipeline runs in one Supabase Edge Function (supabase/functions/chat-proxy) and streams the answer back as it's written.

Step	What happens	Where
Understand	Detects the game (including numbered sequels), the kind of question (build, meta, lore, troubleshooting, speedrun, review, comparison) and whether the question needs fresh data.	index.ts, temporalDetector.ts
Research	Queries several search backends in parallel (Serper, Brave, SearXNG, DuckDuckGo, Google CSE), game wikis via MediaWiki, the Steam News API for patch notes, CheapShark for prices, and official game APIs where they exist. Results are ranked by source authority and recency.	pulseEngine.ts, webSearch.ts, recencyRanker.ts, officialSources.ts
Screenshots	Two passes run in parallel: a dedicated pass reads every piece of text off the image, and a full-resolution crop of the HUD is added so small hotbar details stay legible. The vision model then answers with both.	visionPipeline.ts
Cross-check	Scores how far the sources agree, and flags claims only one source makes.	corroboration.ts
Answer	Ranks models across Gemini, Groq, OpenRouter and Cerebras for each request by cost, health and remaining quota, spending free tiers first and skipping anything rate-limited or failing. Available models are discovered live, so a retired model doesn't silently break routing.	meshRouter.ts, modelCatalog.ts
Shield	Rewrites the instructions around the player's progress, then checks the answer and its suggested follow-ups for anything past it.	spoilerShield.ts, spoilerGuard.ts

Engineering notes

The parts I'd point a reviewer at:

• Resilience on free tiers. Provider health is a circuit breaker stored in Postgres, so every edge-function instance knows a provider is down rather than each one re-discovering it with a timeout. Gemini's free quota is tracked per model, and requests rotate to whichever model has the most headroom left today.
• Rate limiting in the database. One append-only events table answers minute, hour and day windows in a single atomic check. Anonymous visitors are keyed by a salted hash of their IP, never the IP itself.
• Auth that can't be forged. Tokens are signature-verified before a player profile is loaded with the service key; an unverified sub claim would have let anyone read another player's profile.
• Hardened proxies. The wiki proxy only reaches allow-listed hosts (with SSRF tests), and /health never relays a vendor's raw error text.
• A Discord business model with tests. Free, Pro and Premium Server tiers are enforced by a Postgres function. Its SQL runs on a real Postgres (PGlite) in CI, and Stripe handles billing.
• Tested, not just typed. 29 suites and 1,300+ assertions cover game detection, retrieval, spoiler handling, Discord formatting, quotas, billing, share links (including hostile input) and a golden-set retrieval eval. CI runs them, builds the site and syntax-checks the bot on every push.
• A self-audit. AUDIT_REPORT.md records adversarial testing against production: what broke, why, and what was fixed.

Live player count

Under the hero buttons, the landing page shows a live line in the form ● 1,284 players · 9,312 answers researched · 412 games (example numbers), read from gg_public_stats() (supabase/migrations/20261002_public_stats.sql).

• Players: distinct people who got a researched answer, on the web or in Discord (signed-in accounts, Discord users, and anonymous visitors by salted IP hash). It's an honest approximation: one person on two networks counts twice, and a household behind one router counts once.
• Answers: researched answers. Cache hits and stealth turns are never logged, so this is a floor, not a ceiling.
• Games: distinct games those answers were about.

Only these three totals leave the database. The function recomputes at most once every ten minutes, and if it isn't deployed, the line simply doesn't appear. It never shows a made-up number.

Tech stack

Layer	Built with
Web app	React 19, Vite 8, Tailwind CSS 4, GSAP, Framer Motion
Backend	Supabase Edge Functions (Deno, TypeScript), Postgres with RLS
Hosting	Vercel (site + small API routes for wiki and Steam art)
Discord bot	Node.js, discord.js; deploy configs for Fly.io, Railway and Render
Payments	Stripe (Discord Pro and Premium Server plans)
AI	Gemini, Groq, OpenRouter and Cerebras, routed by health and quota

Run it locally

You need Node 22.6+ and Deno. Full walkthrough, including which free API keys matter most: LOCAL_TESTING.md.

npm install
npm install -g deno

cp supabase/functions/.env.example supabase/functions/.env   # model + search keys
cp .env.example .env.local                                   # VITE_SUPABASE_URL=http://127.0.0.1:8000

npm run dev:api   # edge function on http://127.0.0.1:8000
npm run dev       # site on http://localhost:5173

Run the tests:

(cd discord-bot && npm install)   # one suite reads discord.js's permission flags
npm test

Deploy

1. Database: supabase db push applies everything in supabase/migrations (rate limits, provider health, player memory, request trace, public stats). If you run the Discord bot, also run discord-bot/schema.sql and discord-bot/schema-v3.sql in the SQL editor.
2. Backend: supabase functions deploy chat-proxy, then add your keys under Project Settings → Edge Functions → Secrets, plus SUPABASE_JWT_SECRET and RATE_LIMIT_SALT.
3. Site: import the repo into Vercel and set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY. Optional: turn on Vercel Web Analytics and set VITE_ANALYTICS=vercel (cookieless).
4. Discord bot: see discord-bot/README.md.

Repository layout

src/                     React app: landing page (site/), chat (codex/), hooks, utils
api/                     Vercel functions: wiki search/article proxy, Steam art lookup
supabase/functions/      chat-proxy: the whole answer pipeline
supabase/migrations/     Postgres schema (rate limits, health, memory, trace, stats)
discord-bot/             The Discord bot, its schema, billing and Watchtower
tests/                   29 suites, including a golden-set retrieval eval (npm test)

Known limits

• It runs mostly on free model and search tiers. That keeps it free to use, but quality and speed depend on which providers have headroom at the time.
• Retrieval is only as good as what's indexed. Very new or very obscure games get thinner answers, and it says so rather than guessing.
• The player count is an approximation (see above).

Contributing and security

Issues and PRs are welcome; see CONTRIBUTING.md. Please report vulnerabilities privately as described in SECURITY.md.

License

Proprietary. © 2026 Tanmay Angarkar, all rights reserved. See LICENSE.

---

Built by Tanmay Angarkar · LinkedIn · GitHub · [email protected]