Claude Code
Loads SKILL.md as a skill and runs whole discovery cycles from one prompt. Queue five families on your symbol, come back to a leaderboard and the learnings.
"Run a discovery cycle: 5 diverse families on XAUUSD 1h, budget one evening."
From your idea to a strategy you can trust.
No Pine coding. You talk to an AI agent in plain English — the lab does the building, testing and judging, and shows its work at every step.
Quickstart · ≈ 10 minutes
From zero to your first cycle.
Three clicks and one paste — no Pine coding, no backtester to wire up.
1SET UP
Open the kit in your AI agent
Unzip the kit, then open the folder in Claude Code, Codex, Cursor or Gemini — your AI coding agent. This is where you’ll talk to it.
›_any agent that can run shell commands
then double-click START-HERE.bat · start-here.command on Mac
2SET UP
START-HERE opens TradingView for you
That one double-click installs the TradingView MCP — the bridge that lets your agent control TradingView — and connects it automatically, then launches the chart app and health-checks. When it says READY, your agent ↔ TradingView are linked. You set up nothing by hand.
›_Node 16+ · TradingView Desktop (free plan ok)3SET UP
Pick your chart in TradingView
In the TradingView that just opened, load the market you want (e.g. gold, 1h) and leave it. That’s what the lab measures on — you don’t type here, just pick the market.
›_now you’re set up — time to talk to your agent ↓The loop · hands-free after you hit enter
You talk. It does the rest.
Once it’s set up, the whole cycle runs itself: ask → build → test on your real TradingView → keep only what survives.
STEP 01
You ask — in your AI agent
Back in Claude Code / Codex — the same AI agent you opened the kit in — describe an idea or paste a recipe. This is where you talk to it (not in TradingView). No Pine coding.
›_"test a pullback after a breakout on gold, 1h"✦CLAUDE CODE · YOUR AGENTonline
test a pullback after a breakout on gold, 1h
composing 24 variants across 3 families…
↑ you type this in your AI agent — not in TradingView
STEP 02
It builds — many at once, automatically
Hands-free, the agent writes complete Pine v6 strategies from the block library — dozens of real, different variants at once, not one lucky guess. You don't code or wire up anything.
›_18 entries × 8 filters × 9 exits = 3,888 on tap◳ BUILD QUEUE3,888 combos
donchian-breakout-021✓ built
supertrend-flip-009✓ built
fvg-retrace-016✓ built
gap-play-004✓ built
nr7-break-012◷ testing
STEP 03
It tests on your real TradingView
Every variant runs in your own Strategy Tester with real fees, across many windows — about two seconds each, fully automatic; you don't click through anything. Signals never repaint (never change after a bar closes), so what you see is what you could have traded.
›_real bars · real costs · no simulated candlestradingview.com · Strategy Tester · XAUUSD 4hreal run
STEP 04
You get a verdict — and the source
Each one comes back PASS, WEAK or KILL with the reason on file. Most die — that's the proof it's honest. Survivors arrive as full Pine source you keep.
›_PASS · WEAK · INSUFFICIENT · KILL — cause loggeddonchian-breakout-021XAUUSD · 1D
PASS
OOS PF 3.35retention 217%t 4.76
55 survived of 251 · the rest on file
You watch the whole thing happen — every build, test and kill is a plain file you can re-open and re-run. Nothing is hidden, and the best a strategy ever earns here is “promising — validate further,” never “profitable.”
After the verdict — it all lands on your dashboard.Every survivor, kill and dossier files itself into Mission Control: the network, leaderboard, Edge Map, and copy-paste Pine. See exactly what the agent produces.
See the dashboard & proof →the real 251-candidate board
A real cycle
Your agent types. The gates decide.
One prompt — to Claude Code or any agent that can run shell commands — and the factory builds, lints, measures on live market data, and files the verdict. No simulated candles. No homebrew backtester to argue with.
- 100+ candidates built and judged in a single evening session
- Real commission and slippage compiled into every strategy
- Non-repaint by construction — GATE-0 lints it before a single test
- Every verdict re-runnable offline from results.json
node tools/build-strategy.mjs --spec st-flip.yaml --out lab/candidates/st-flip-014
✓ strategy.pine composed — SuperTrend flip × session filter × ST trail · long/short
✓ GATE-0 offline lint — non-repaint ✓ · real costs ✓ · window inputs ✓
node tools/validate.mjs lab/candidates/st-flip-014/strategy.pine --symbol XAUUSD --tf 1h
year sweep → chart carries 2019–2026 · sized IS 2019–23 · OOS 2024–26
window 1/6 · PF 1.91 · 198 trades · maxDD 5.8% [2.1s]
window 4/6 · PF 1.55 · 207 trades · maxDD 7.4% [1.9s]
window 6/6 · PF 1.62 · 214 trades · maxDD 7.0% [2.0s]
node tools/verdict.mjs lab/candidates/st-flip-014
VERDICT: PASS — “promising, validate further” · gates v2.1 · trials 1
node tools/library.mjs regen --lab lab
leaderboard.md ✓ · graveyard.md ✓ · learnings.md +2 lines
Under the hood
A strategy factory with a kill floor.
Generate → build → validate → judge → remember → evolve — hands-free at ~2s a window. Volume from the factory, truth from the gates.
↺ Before every cycle the lab reads its MEMORY — learnings.md · graveyard.md · leaderboard.md — that’s how it gets smarter
01GENERATEDiverse hypotheses — popularity gets tested, never believed.
02BUILDComplete Pine v6 from a one-screen spec — loop-ready by construction.
03GATE-0Offline lint — non-repaint, real costs, window inputs. Free, in 5 seconds.
04MEASUREYour real Strategy Tester, hands-free, ~2s a window.REAL TV
05JUDGEKill-gates → PASS / WEAK / INSUFFICIENT / KILL.
06REMEMBERLeaderboard, graveyard, learnings — plain auditable files.
07EVOLVE≤2-param mutations, plateaus not spikes, snooping made visible.
Windows the data can’t answer print as ? — the loop is structurally incapable of the classic “great results on data that wasn’t there” lie.
No lock-in
Claude Code, any agent, or just you.
The tools drive TradingView through a plain CLI, so MCP support is optional — and every stage can be run by hand.
Cursor · Codex · Gemini CLI
Point any command-running agent at START_HERE_AI.md + AGENTS.md — plain instructions, plain CLI. If it can run shell commands, it can run the lab.
node tools/validate.mjs … && node tools/verdict.mjs …
Manual mode
The complete loop as a human checklist. You propose candidates and write the learnings — every measurement, gate, dossier and dashboard is identical.
docs/MANUAL_MODE.md — same tools, same gates, same honesty.
After a PASS — now what?
From a passing strategy to a buzz on your phone.
A PASS is a hypothesis, not a green light. The lab maps the honest path — and it’s 100% within TradingView’s rules.
1
default — before money
VALIDATE MORE
Forward-test it live with the kill-line watchdog; replicate it on a market it's never seen. An edge on one tape is usually noise.
2
make it sturdier
EXTEND / IMPROVE
Mutate carefully, stress-replay it, and build a structurally-independent portfolio — let the real correlation matrix decide, not a clone.
3
ToS-clean alerts
GO LIVE
Each survivor ships a webhook-ready alert Pine + risk card + a ~60-line Telegram relay. Official TradingView alerts — and it never places an order.
Official TradingView alerts + webhooks — the ToS-clean path. The ~60-line Telegram relay relays notifications only and never places an order.
It’s all yours
Full source. Yours to make better.
No black box, no lock-in. You own every line — fork it, add your own blocks, tighten the gates, and build something far beyond what we shipped.
⌘
The complete source
All 15 tools, the kill-gates, the builder and the Pine it writes — every line editable and yours to extend privately. We hand you the whole quant spine, not a locked app.
❋
And the Mission Control dashboard
The whole lab as one living board — THE NETWORK, leaderboard, Edge Map, fund-style dossiers with copy-paste Pine, a Monte-Carlo Sim Lab, portfolio correlations and the expectancy surface. One self-contained file, ships in the box.
See the dashboard →Ready to run your own ideas?The full 251-candidate demo lab ships in the box to learn from first.
Get the lab — $149 $79one-time · runs tonight