thornberry
Claim your seat
thornberry ai packs the things a real desk runs on into one local app: a research crew, a quant lab, a global macro monitor, and a risk engine. The AI does the reading. A strategy builder and backtester turn your ideas into rules you can trust. Native engines do the heavy math. Your bots place the trades, inside your limits. Mac now; Windows and Linux soon.
At the core, three jobs that never share hands: the AI reads, your strategy decides, your bots trade. Wrapped around that: a strategy builder and a backtest lab, native math engines, an eight-feed global monitor, a vault that teaches the agents how you think, and a remote you can run from your phone. All of it local. All of it yours.
A Captain coordinates; equity/futures/options Commanders rank candidates; Scanners, Technical, Macro, Sentiment, Flow, an Oracle live-web agent and more feed context up the chain.
The execution layer runs your defined entry, exit, sizing, and risk rules against the AI's candidates. The AI proposes; your strategy disposes.
It holds your limits — position size, drawdown, prop-firm rules — and owns the kill switch. The bots trade inside those limits or not at all.
Paper trading runs against live data, on its own clock. The backtester runs up to 5 concurrent walk-forward + Monte Carlo backtests.
Graduation tiers. A strategy climbs the sizing ladder only after 15+ calibrated trades per market regime, with deflated Sharpe and drawdown bounds respected.
It holds your style, rules, and journal as typed context the agents read directly. Real context, not prompt-stuffing.
Local models — any size your hardware handles — do the grind; frontier models take the hard calls; risk-path agents stay on verified ones. Your call, per agent.
Pair Telegram, Discord, Slack, WhatsApp, Signal, or iMessage to ask questions, journal, and nudge the crew. Trades stay in the app — a message can't place one.
Streaks, P&L, crew health, the morning briefing — turned into something you want to check, not a dashboard you dread.
No canned strategies ship with thornberry ai. You bring your own. Build one in a real editor (or have an agent draft it from your vault), backtest it hard, paper-trade it, and let it earn size as it proves out. Paper-first is the default; live is your call, and so is the moment you flip it.
A typed editor where a setup can demand real confluence ("three of these five have to agree") before it'll even look at an entry. Or let the strategy-architect draft one from your own notes.
Walk-forward and Monte Carlo on the actual logic you'll run, not a simplified stand-in. Sharpe, Sortino, Calmar, drawdown, expectancy.
Every new strategy starts on paper against live data, on its own clock. Edit it for real and the clock resets, because it's a different strategy now.
It climbs the sizing ladder as it builds real history across different conditions, and steps back down when a drawdown or a cold streak says it should.
Flip live when you mean to. Your bots run the strategy inside your limits; at L1, every live order still stops at you first.
You never gave anything your keys. You never let a model pull the trigger. The desk runs when you do, your strategy decides, your bots execute, and your limits hold. That's the day.
Every step — the candidate, the rule that fired, the risk check, the fill — is written to an encrypted log only you can read.
At L1, the order comes to you first: here it is, here's why, approve or pass. Every new strategy starts in a paper window, so by default it fills on paper and your real account never moves. Once a strategy has earned live and you've turned live on, the same approved order goes to your broker — and you still approve each one.
— position cap, drawdown rule, the strategy's tier, the kill-switch state. Anything outside the lines gets blocked before it leaves your machine.
Your strategy builds the order — entry, size, stop, target, all from your rules — and the bot is ready to place it.
The execution layer runs the rules you wrote against the ranked setups. One needs three of five confirmations to align before it'll consider an entry. Two align. It waits.
Scanners screen the universe down to a short list. The technical, macro, sentiment, and flow agents annotate what's moving and why. The world monitor flags an overnight seismic event near a shipping lane and a jump in a prediction market. None of it is advice — it's a ranked, sourced briefing in your overview.
You open the desk and it gets to work — thornberry ai runs while you're in front of it, not behind your back.
One giant model bluffing its way through every job is how you get confidently-wrong calls. thornberry ai runs a crew instead, each one good at exactly one thing, arguing it out up the chain. A Captain keeps them pointed the same direction. Commanders for equities, futures, and options turn the noise into a ranked short list, each leaning on a local model for the grind and a frontier model only when the call is genuinely hard. And you can add your own. Not one of them can place a trade.
A Captain, three Commanders, their lieutenants, and strategists.
Scanners that comb the whole universe down to a handful worth your time.
Technicals, macro, sentiment, a live-web Oracle, news, calendar, earnings, sector rotation, options flow, correlation, patterns, quant.
Multi-timeframe chart context, cleaned up per asset class.
Recon, heartbeat, watchlist, triage, and the world monitor keeping the lights on.
A backtest orchestrator and per-asset position trackers.
The risk engine, off in its own corner, holding the limits.
And the whole team reads from your vault: your rules, your journal, your style, as real typed context. The longer you use it, the more it researches like you would, because it's reading what you actually think, not guessing.
The number-crunching runs natively in Rust, on your machine. Fast, repeatable, and nothing leaves the desk.
RSI, MACD, Bollinger, ATR, ADX, Stochastic, CCI, Williams %R, Hurst, realized vol. All recomputed every bar, instantly.
Trend state, swings, Fibonacci, order blocks, supply/demand zones, and the reversals every chartist squints for: pin bars, engulfings, head-and-shoulders.
How your book moves together: pairwise correlation, beta, net-beta, historical VaR, and a read on which regime you're actually in.
Sharpe, Sortino, Calmar, Kelly, expectancy, recovery factor, plus Monte Carlo and stress tests that try to break your strategy before the market does.
Eight live global feeds, polled on your machine and pinned to a 3D globe. The context most trading tools pretend doesn't exist.
Per-feed alerts and geo-pinned events, sourced from USGS, Polymarket, GDELT, CISA, and more. An earthquake near a shipping lane is a chart move three days early, if you're looking.
Click into any of them. Here's the vibe.
Overview
Command Center
Agent ControlsLive TradingPortfolioAnalytics
Research
Alpha IntelBacktest
Chat
World MonitorSame data, two completely different moods. Flip whenever.
Full neon. Glow, motion, a little witchcraft, the markets looking back at you. The default, and the one you'll screenshot.
Lights off. Flat, quiet, high-contrast, for deep focus, tired eyes, or a screen reader. Fully accessible at launch.
thornberry ai is a platform, not a sealed box. Write your own research agents, teach the crew new skills, and add brokers through the plugin SDK. Whatever you build plays by the same rule: the AI researches, your strategy decides, your bots trade inside your limits.
A role prompt, a model tier (local or frontier), and it slots into the crew. It researches and reports like the built-ins. Like them, it never trades.
Teach the desk a new move: a feed, an indicator, a screen, a morning ritual. Typed, local, no cloud round-trip, no leaking your edge.
New broker connections through the SDK; new chat channels through pairing. Bring thornberry ai to the stack you already run.
Your rules, your journal, your style live in a local vault every agent reads. The more you put in, the more it researches like you, because it's reading what you actually think.
The platform is open; the guardrails aren't. The AI researches, your strategy decides, your bots trade inside your limits, whatever you bolt on.
Apple Silicon (M1 and later) and Intel, with Metal-accelerated local inference. The home turf. Notarized & stapled.
Windows 10 22H2+ / 11, x86-64 with AVX2. WebView2 bundled. EV-signed installer.
Ubuntu 24.04 LTS / Fedora 39+ via AppImage with a detached signature.
A normal laptop with 16GB of RAM is plenty. Pick a local model that fits your machine: small if it's small, big if you've got the GPU.
⚠ AI and your security. thornberry ai runs AI research agents on your own machine. AI can be wrong: it can misread the market, surface flawed or incomplete research, and make mistakes. Treat every output as a starting point you review yourself, never a recommendation. You write the strategy, you set the limits, and your bots trade only inside them, so every order, and the security of your machine, keys, and broker connections, is your responsibility. Keep new strategies on paper until you trust them. thornberry ai is software, not advice, and is not responsible for AI errors, trading losses, or any security issue on your device or accounts.
thornberry ai is trading software, not a broker, and not investment advice. Trading securities, options, and futures carries a substantial risk of loss and is not suitable for everyone. You can lose some or all of your capital, and with options and futures you can lose more than you deposit. Past performance does not predict future results. You make every trading decision and you bear every result.
