Install a plugin in one line. Ask in plain English. Chain plugins into flows that produce a film, a trade report or a printed book.
Four building blocks. Each one builds on the one before it.
A toolbox that does one job, like market data or video.
price@hunzai-agentsThe actions a plugin adds. Type a slash command or just ask.
/sol is SOL at a high?Helpers that take on a longer job in their own space.
leverage-researcherSeveral plugins chained into one job, with checks along the way.
research → film → uploadReady-made pipelines. You type one sentence; the plugins run in order and stop for your OK before anything costs money.
Get a scored long/short verdict with entry, target and stop — and a phone alert when price gets there.
Price moves, news and the week’s calendar in one image card on your phone.
Researched, written and filmed part by part. You approve each prompt before anything is generated.
A hand-drawn explainer video, voiced in other languages and ready for YouTube.
Chapters with matching illustrations, laid out as a printable PDF.
Each one does one job and installs on its own.
Actually place the trade — spot swaps and leveraged perps on Solana.
Jupiter — spot swaps and leveraged perpetuals (SOL/BTC/ETH) with take-profit, stop-loss, and limit orders. Wraps the official `jup` CLI.
Market data for everything else to trade on — one snapshot, every feed.
The canonical market-data plugin — ONE source-tagged `snapshot` JSON (single or multi-symbol: --symbols sol,btc) with every live feed: real-time Pyth prices, RSI/MACD/Bollinger/EMA signals, pivot levels with touch-count most-tested S/R (1h–24h), a computed liquidation-tier heatmap flagged against real S/R, Fear & Greed sentiment, market breadth, reversal diagnostics (RSI & CVD divergence, Stochastic RSI, Bollinger squeeze, OI↔price), OKX + Binance derivatives flow (funding, OI, long/short, liquidations), macro quotes (WTI/gold/DXY/US10Y via CNBC) and registry-driven RSS news.
Research a leveraged entry before you take it — and backtest the idea behind it.
Pre-trade leverage research + deterministic backtesting. The leverage-researcher agent grades its own past theses first, runs a 9-dimension scored analysis (macro gate → structure → momentum → funding → verdict, RR gates ≥1.5R) from one price-plugin snapshot, and persists every verdict as a thesis file with machine-readable alert triggers (watched by the watcher plugin); the `vt` CLI backtests strategies with Monte Carlo + bootstrap validation — pure Python, zero LLM. Research only, never trades.
Turn a "wait for X" thesis into an alert that actually reaches you.
Deterministic trigger monitor — polls the thesis files written by the vibe-trading leverage skill, evaluates price/1H-close triggers via the price plugin, and sends a Telegram alert the moment one fires. The standing watch that turns 'WAIT + invalidation conditions' into an actual signal. No LLM, never trades.
Ask what-if of a market you already recorded.
Market recorder + hypothesis replayer for SOL — records live price/indicators/touch-clusters into one CSV per day (10s cadence), then replays recorded days under what-if hypotheses (limit/market/cluster/consolidation entries, leverage, TP/SL, time exits) and sweeps parameter grids with a real fee model. Research instrument: it never trades. Composes the price plugin.
Learn which minutes are tradeable after fees — and which are noise.
Fee-anchored regime classifier for the scalper — trains a 3-class up/down/flat model on 1-min candles where 'flat' means nothing tradeable after fees, evaluates with purged walk-forward only, and gates promotion behind GO/NO-GO thresholds (precision lift, detection lead, flap rate).
Babysit an open position so the TP/SL actually fires.
Deterministic TP/SL babysitter for open Jupiter perp positions — polls open positions, closes on fixed take-profit/stop-loss thresholds, and brackets naked fills with TP/SL triggers. DRY-RUN by default; no AI in the loop.
One card on your phone that says what the market is doing, and why.
Market-analyst engine — compute + card render only (all data comes from the price plugin). ONE combined Telegram card (top reversal-strength banner STRONG/MILD/WEAK over MARKET / NEWS / CALENDAR sections) and a 3-part intelligence JSON (/intel: geopolitical news, BTC+SOL leverage indicators, macro oil/gold/DXY/10Y — all via price). Composes the price + telegram plugins; ships the market-analyst agent.
Get results out of Claude Code and onto your phone.
Telegram Bot messaging — send text, photos, documents, audio, and video
Read, search and answer mail without leaving the terminal.
Gmail API — read inbox, search, send, and reply to emails via OAuth2
Publish a finished film or song to your channel without filling in YouTube Studio by hand.
YouTube Data API uploader — reads a social package folder (title, description, tags, category, thumbnail, AI-content and made-for-kids answers), shows the plan, then does a resumable, private-by-default upload.
Ask the live web a question and get sourced answers back.
Perplexity API — live web search (/search), agentic answers with end-to-end function calling (/ask), and text embeddings (/embed).
Turn a topic and a pile of sources into one clean research document.
Topic → one well-structured research.md. Researches a topic (or question/brief) into clean markdown, optionally grounded in supplied sources — URLs (crawled via firecrawl), local files/folders, PDF/DOCX, or raw text. The front half of the illustration pipeline.
Turn anything into a shareable card image — no browser, no API.
Generic declarative card renderer — a JSON spec (sections, tiles, grids, events, badges, bars, kv tables, images, columns) → a crisp PNG via pure-JS satori + resvg (no browser, no API, no LLM). Turn ANY input into a card image in one shot.
Turn a document into a chaptered, colour-consistent picture set.
Document → chaptered, colour-consistent illustrations. compose reads an input document, chapters it, and authors palette-consistent prompt JSONs (default colorful pencil-sketch); render turns any prompt-JSON folder into images (seedream-5-lite ↔ nano-banana-2 fallback); stylize caches hosted URLs for image-to-image consistency.
Give a still image motion that obeys real-world physics.
Replicate image-to-video renderer (ByteDance Seedance) that directs like a videographer — reads the still, picks the shot and one of 8 official camera moves, and writes physics-obeying motion prompts. Preserves real photos as the first frame, adding only motion. Folders of video-prompt JSONs → MP4 clips.
Split research into narrated documentary parts you can film one at a time.
Splits a research document into an ordered set of documentary PARTS — one narrated shot each — writing outline.md, a global style.md, and a self-contained NN_<part>/ folder per part (script.md, timeline.json, references). Separates documented fact from labelled reconstruction; writes no media itself.
Cut the shots, voiceover and music into a finished film.
Documentary timeline.json → a finished film via ffmpeg — Ken Burns on stills, shots normalized to one aspect/resolution/fps, voiceover+SFX mixed per shot, optional ducked music bed, burned captions and title cards; --join concatenates per-part videos into the final film. Pure post-production, generates no media.
Compile chapters and images into a printable book PDF.
Book PDF compiler — folders of NN_<slug>/chapter.md markdown (with per-chapter images) → a typeset A4 PDF via pandoc + tectonic. Bilingual columns, two-column, facing-plates and half-page picture-book layouts; RTL/Urdu auto-detected; themeable via book.yaml.
Find out which skills a field actually pays for, from real postings.
Skills-demand engine — deterministic CLI that normalizes, deduplicates, aggregates and snapshots AI/blockchain job postings into a ranked, accumulating skills-demand database.
Drive any website like a human — visually, not through its API.
Screenshot-driven, step-based agentic browser loop via playwright-cli. Opens a dedicated isolated Chrome profile, plans the goal into small verifiable steps, and works each with up to 3 tries — persisting before/after evidence and a full run ledger to disk. Visual + DOM only; never the site's API or app internals.
From a question to a finished, narrated film — script, animation, voice, sting, social pack, dubs.
Research a question into a script of question-led chapters, then a narrated hand-drawn motion-graphic film: a chapter card per chapter, one labelled moving shot per narrated sentence with its point on screen, one ElevenLabs narration, and film-full.mp4 that opens with the Ibex Labs sting and a 2 s education-and-entertainment-only notice and closes with a research-sources end card. Plus a fact-checked social-media package and reviewed translations into other languages.
Write a book for the ear and narrate it chapter by chapter.
Audio-native book writer — author drafts chaptered sections written for listening (never adapted from print), a lint gate enforces the book spec before any credits are spent, narrate voices one MP3 per chapter with ElevenLabs v3, and publish ships the mastered audio to GitHub.
Helpers for longer jobs. They work in their own space so your main chat stays clean.