How Coil works

The short answer: Coil is an agent-native trading copilot you run on your own machine — Mac or Windows — inside your own AI agent (built for Claude / Claude Code), with your own broker and your own capital.

It does three things in one: a scanner scores every name in the Equities book (S&P 500, Nasdaq-100, and a Macro book of bonds, income, gold and metals, commodities — ~560 names); a dashboard shows you the whole board — the picks, the posture, the sector rotation, the "why" behind each score; and a long-only engine trades those published scores by rule. It buys leaders at real entries, never chases what has already run, never catches a falling knife, and sits in cash when nothing qualifies. It ships disarmed, and nothing here is a signal service, managed money, or investment advice. Below is exactly how each piece works, and where it's weak.

The thesis: buy the leaders, at the moment they're ready

Markets are led by a rotating handful of names. In one regime it's semis; in another it's financials, energy, or health care; sometimes leadership leaves equities entirely and the safe money is in bonds or gold. The durable edge isn't picking a sector and marrying it — it's owning whatever is actually leading, and only entering when the entry is real. That's the whole idea in the name: Coil looks for a leader coiled at support — compressed, holding, ready — and strikes when it releases, rather than piling in after the move is obvious. It just does that across the entire market now, not one corner of it.

The non-obvious part is the discipline, not the ambition. Anyone can hold a list of strong stocks. The hard, unglamorous work is refusing the two entries that quietly lose money: buying a name that is still falling (it looks cheap; it keeps getting cheaper) and chasing a name that has already gone vertical (it looks strong; you're the last buyer). Coil is built to say no to both.

1. It scans and scores every name

Every cycle, the scanner walks the full Equities book — the S&P 500, the Nasdaq-100, and the Macro book — and scores each name on a set of independent, readable factors. No black box; the dashboard shows you the number and the research note behind it. The factors that matter most:

Those factors combine into a single opportunity score per name, and the scanner also maps a name to a leveraged vehicle where one exists — NVDA to NVDL, the Nasdaq-100 to TQQQ, semis to their 3x ETF — so the engine can accelerate a high-conviction leader at a fraction of the notional. The dashboard is the source of truth: it publishes the scores, the Top-6 picks, the posture header, the sector-rotation view, and the structural stop and target for each name. The engine never re-guesses the board; it trades what the scanner publishes.

2. The entry discipline: READY vs CHASE vs FALLING

This is the part most tools skip, so it gets its own heading. A "strong stock" is not a signal — where it is in its move is the signal. Coil's entry window is a hard gate, and it exists to keep you from the two losing entries:

• READY / SETUP — buy. A leader that has pulled back into support, compressed, and is holding or turning up at a real, structural entry. This is the coil. Volume-profile and Fibonacci support/resistance define the level; the engine wants to be here before the move, not after.
• WAIT — not yet. A good name, but not at an entry. The engine watches and waits for it to come to the level rather than paying up for it today.
• CHASE — refuse. A name pinned at its highs, extended after a thrust. It may keep going; it may also hand back the whole move to the first buyer at support. Coil will not chase it — an intraday-thrust guard specifically downgrades names sitting on today's highs so a rec is truly ready, not chasing the open.
• FALLING — refuse. A name still in decline. "Cheap" is not an entry; a falling knife keeps falling. The engine buys weakness inside a confirmed uptrend, not weakness that is still breaking down.

The edge Coil is built on is buying disciplined weakness in an uptrend, not chasing breakouts and not bottom-fishing. Standing down is a position — on most cycles, most of the board is WAIT, CHASE, or FALLING, and that's intentional.

3. The engine: rank, size, accelerate, exit

The engine reads the published board and acts by rule. It is long-only by design — there is no short book to invert, no leverage you didn't opt into.

• Rank by buy score. Among names that are actually at an entry, the engine ranks by a composite buy score (opportunity, leadership, entry quality, posture) and works down the list — the best available signal wins, wherever it sits in the Equities book — the S&P, the Nasdaq-100, or the Macro book. One unified pool, no per-book quota.
• Size by conviction. Cleaner, stronger setups earn more size; marginal ones earn less. Every size the strategy asks for is then clamped by the safety spine (below).
• Accelerate with leverage, at reduced notional. For the highest-conviction leaders, the engine can express the position through the mapped leveraged ETF — but sized down, so the leveraged exposure is a fraction of a full position, not a full position tripled. Leverage is an accelerator on the best names, never the default vehicle. (Why that matters, and why leverage decays, is its own topic — see leveraged-ETF decay.)
• Exit on a ladder, trail the rest. Positions scale out at structural targets and carry a trailing stop that follows price in your favor. The stop is the backstop, not the plan; the primary exits are rule-based, taken when the reason for the position stops working.
• Cash is a position. When the pool is thin — few leaders at real entries, defensive posture — the engine raises cash rather than force a trade. Idle cash sits in the broker's own cash sweep (variable, set by the broker, e.g. Robinhood Gold quoted ~3.35% APY in early 2026 — the broker's yield, not Coil's, and not risk-free). In a genuine down market, "down" means more cash and a defensive rotation into the Macro book — bonds, income, gold — never a short.

The honest caveat on exits: a stop is not a guarantee. Markets gap, and a position can open beyond its stop — a stop cannot fill inside a gap. That risk is real and permanent, and it is exactly why the engine sizes leveraged vehicles down.

4. The safety spine — the failsafe, not the steering

Everything above manages opportunity. A separate, deliberately dumb layer sits underneath it and bounds damage. The strategy proposes; the safety spine disposes. It is the seatbelt, not the steering wheel:

• Account whitelist. Coil can only touch the one account you explicitly allowlist. Every other account you hold is off-limits by construction.
• Hard caps. A ceiling on any single name's share of equity, and on total exposure — so the book is never fully concentrated in one position, even at maximum conviction.
• Drawdown ladder. A layer that watches your loss on the day and your cumulative drawdown from your equity high-water mark, and escalates as it deepens — halt new entries, then cut size, then cut further, then a full stand-down. Read it as an account-level limit across a bad streak, not a per-trade loss; a single trade is cut at its own exit long before any of this.
• Kill switch, and it fails closed. One command halts everything. And if anything is ambiguous — a data gap, a failed check, an integrity mismatch — the engine does not trade. It fails closed, toward doing nothing, never toward guessing.
• Ships disarmed. Out of the box LIVE_TRADING is off. You dry-run first, and going live is a separate, deliberately gated flow: allowlist your account, re-pin the integrity fingerprint, pass the safety self-test, and type a total-loss acknowledgment. Live orders route over Robinhood's official agentic-trading rails — dedicated agentic accounts connected via MCP — with this safety spine layered on top, because Robinhood provides the rails but "does not control, supervise, monitor, recommend, or audit these AI agents."

Plainly: the spine reduces and bounds damage; it does not remove market risk. Leveraged ETFs, where the engine uses them, can lose value rapidly, including total loss of the amount committed to them.

5. The research backbone — and the honesty rider

Numbers are only as honest as the machine that produces them, and the frame around them. Coil's scoring is built on a research backtest, not a live track record: the leadership-rotation backbone replayed on point-in-time index membership (delisted names included, so there's no survivorship bias), with next-open fills and trading costs modeled, from 2017 through the first half of 2026. Here is the whole record, with SPY next to every number:

Cumulatively the backbone compounded +638% versus SPY's +282%, with a shallower worst drawdown (−23% vs −32%) and a positive result in 9 of 10 years — the only red year was 2018, at −1%, when SPY fell −7%.

Self-tuning, on a leash

Coil adapts, but only where adapting is safe. The validated rule-set is frozen and fingerprinted — Coil is not an open-ended AI inventing new trades on your machine. What adapts is deliberately narrow: position sizing, and only sizing. Every close, a local learning pass (python3 -m coil_v4.learn) reads your own closed round trips and leans size toward what has been working for you — nothing else moves.

• What it tilts. Sizing by book (S&P 500 / Nasdaq-100 / Macro) and by buy-score tier. It never touches which names it buys, the entry windows, the exits, the stops, or the caps.
• The clamps. Each tilt is clamped, and the stacked product is clamped again to roughly 0.55×–1.35× overall. Tilts are shrinkage-weighted, so one trade barely moves the dial — it takes a history to move much — and the loop is a neutral no-op until you have about two or more closed round trips behind it.
• Live only. It learns from live round trips only; the ledger is mode-isolated, so nothing tunes until you arm the engine and close real trades. Dry-runs and backtests never feed it.
• The audit trail. Every tilt is written to a readable log (optimization_log.json) — exactly what changed, by how much, and why. It also surfaces diagnostics it never auto-acts on: which sectors, sleeves, and entry-quality bands have been paying or bleeding.

It runs entirely on your machine — nothing leaves it, and your history persists across packet updates. New validated structure arrives only as a deliberate new version you choose to download; updates are announced on coil.trade and installed manually, never pushed to your machine. The packet ships with the research backtest harness and a safety self-test suite, so you can re-run the research and verify the spine yourself rather than take either on faith.

Where to go next

If you want the operator side — how a scheduled Claude agent and a broker connector actually arm and run this on your machine — read the agentic trading guide. If you're weighing Coil against bots and signal services, the comparison lays out the differences. The whole system is a $29 one-time download, yours to keep and run forever.