Private beta · paper trading
A team of Claude agents reads the market, writes the thesis, proposes the trade and argues the other side. One deterministic program, the Executor, is the only thing that can touch the broker, inside limits you set. Nothing moves without your line.
How it thinks
Bellwether reads the market the way a professional desk does: the overnight news, the filings, the levels, the tape. It forms a view with conviction, writes down what would prove it wrong, and argues with itself before anything reaches you. When the facts change it changes its mind, and says so. It acts only inside the line you draw. Underneath, time-tested, published market signals are fused with our own desk intelligence: the process is what we build, not a secret indicator.
How it worksEvery idea arrives with a thesis, a size, a stop and the price that proves it wrong.
A Critic reviews every session and every new position; objections are answered in writing, not waved away.
Above your approval threshold nothing moves without you, and Halt is one click from every screen.
Illustrative session · fictional symbol and prices
An illustrative session: the Researcher flags a name, the Portfolio Manager writes a thesis with a stop, a target and what would prove it wrong, the Critic objects and is answered, the Executor runs its checks, and the approved order is parked for the open.
The tape: Prints drifting past: direction and pace, never a quote to act on.
On the floor
The tape, the sectors, the levels it drew before it acted, and the names drifting toward a trigger. Rendered live on this page; the symbols are made up and no prices are shown.
Illustrative · fictional symbols · no prices
The team
Six Claude agents with narrow jobs, and one plain program with the keys. They are kept deliberately apart: Claude proposes, code disposes.
Reads the tape, the filings and the news before the open and at midday; writes a cited brief on what matters for your universe.
Never proposes a trade or sees your positions’ dollar value.
Turns research into proposals the way a desk writes a ticket: thesis, evidence, size against your limits, a stop, an invalidation level and a horizon. Keeps a journal.
Never sends an order. It can only ask.
Turns the PM’s intent into concrete order instructions: the limit, what to do if the open gaps, when to scale out. Today a deterministic step inside the PM’s run; a dedicated agent scored on execution quality is next.
Never opens a new name or moves a target, and never talks to the broker.
Reviews every session and pre-checks each new position: sizing, correlation, stale evidence, thesis drift. Its findings must be answered at the next session.
Never approves anything. It can only object, in writing.
Runs the algorithm blend and the backtests, attributes what worked, and recommends bounded weight changes with the evidence attached.
Never applies a change silently: every recommendation is a diff you accept, revert or ignore.
Answers your questions about the book in plain English and turns requests into intent cards you confirm.
Never has a trade tool. A request that implies a trade goes through the PM, the Critic and the Executor like any other idea.
Ordinary, tested code with no model inside. Holds the broker keys, runs every proposal through the same risk checks, holds anything over your approval line for you, then places one day limit order.
Never argues, never improvises, never skips a check. It cannot be talked out of your limits, by the agents or by you.
The rhythm
The same beats every trading day, Chicago time. You can watch all of it or none of it; the line you set holds either way.
The Researcher reads overnight news, filings and levels for your universe and writes a short, cited brief.
The PM proposes: a thesis with the evidence it rests on, a size, a stop and the price that proves it wrong. The Critic attaches its objections.
Anything above the approval threshold you chose waits for you as a card. Below it, your autonomy setting decides. Nothing else moves.
The Executor waits out the first minutes, then re-checks price, size, exposure, loss and drawdown limits at the moment of sending, places a day limit order, and manages the stop it was given.
Every hour the PM looks again at what it is watching: levels reached, theses drifting, stops to tighten. Most hours it does nothing, and writes down why.
A last pass twelve minutes before the bell: anything that should not be held overnight is proposed for sale, inside the same line and the same checks.
The session is journaled, proposals are scored against what actually happened, and the findings are on the PM’s desk tomorrow morning.
Lessons are kept or retired, the Critic’s calibration is re-checked, and tomorrow’s brief starts from what today taught. You are asleep; the line still holds.
Learning
After every close the desk reviews its own decisions: proposals are scored against what actually happened, lessons are kept or retired, and every change to how it trades is versioned, explained and revertible. The evidence is yours to read: the decision journal and calibration today, adherence and a falling count of repeated mistakes as the Learning tab fills in. Recursive self-improvement, with a paper trail, and you holding the pen.
Calls, scored after the close
Illustrative
Calibration
stated confidence against how often it was right; the line pulls toward the diagonal as calls are scored
Lesson
retiredGuidance raises that open on one block print tend to fade by noon: start small, add only after the first hour holds.
retired · the evidence stopped supporting it
The console
Real screens from the console, captured from the running product with sample data, not marketing renders.
Model-written prose is always marked as such, every number is the platform’s own, and anything you might act on links to where it came from.
Built in, not bolted on
The rules that matter are not instructions to a model. They are code paths that do not exist, limits enforced where the order is sent, and a switch that is always one click away.
Every book starts on a paper account: real prices, pretend money. Going live is a deliberate, multi-step change made outside the console after weeks of paper evidence. Never a toggle.
No agent has broker credentials, a network tool or a way to write an order. The one component that can trade has no model in it and is covered by tests line for line.
Daily loss, drawdown, exposure, order size, stale quotes, no trading around the open and close: checked at send time, every time. Loosening a limit asks for your authenticator code and leaves an audit line. Halt is one click from any screen.
Each book is isolated at the database layer: another operator’s agents, data and settings cannot read or touch yours, and a read-only guest sees numbers, never controls.
There is no code that can deposit, withdraw or link a bank. Funding happens at your broker, by you. Bellwether never sees bank details.
The checks
Every order, every time: the Executor’s checks in the order they run. The first failure wins and nothing is skipped.
Evidence
Every strategy ships with an evidence page, and the page shows the misses. Here is one: a published pivot method, implemented literally and replayed through the production risk policy.
n = 3,360 fills over 20.7 years (2006–2026), 28 symbols, 5,453 daily sessions; next-open fills with modeled spread and impact; 230 orders rejected and 124 halt days imposed by the same risk policy the live Executor runs.
It lagged simply holding the index through the same limits. We published that, kept the method as one vote among several, and wrote down the question it raises for its author.
FAQ
A small desk of Claude agents that research, propose and review trades in US stocks and ETFs, one deterministic program that alone can place an order inside limits you set, and a console where you approve, steer and read everything they wrote.
Not by default. Every book runs on a paper account with real prices. Going live requires weeks of paper evidence and a deliberate change made at deployment time, outside the console; if the book and the deployment ever disagree about which world they are in, the Executor halts rather than guess.
The line: an approval threshold above which nothing moves without you, an autonomy level below it, a risk posture, hard limits that only move within code-defined bounds, plain-English guidance the PM must acknowledge, and a Halt button that stops new orders within seconds from any screen.
Three ways, all published in the console. Rules and the risk policy are replayed walk-forward over twenty years of daily data with costs, and the evidence page lists what the replay does not cover. Agent sessions are run on trading days after the model’s training cut-off and scored against what happened, including how well stated confidence was calibrated. And every live paper day is journaled, reconciled to the cent and reviewed by the Critic. None of this predicts anything; it tells you honestly what happened.
Every session is journaled and every proposal is later scored against what the market actually did. After the close the desk reviews itself: lessons that keep earning their place are kept, the rest are retired, and the Critic’s track record is checked for calibration: when it says 70%, is it right about 70% of the time? Changes to how the desk trades are never silent: each one is versioned, explained, and revertible by you. A dedicated Learning tab that charts adherence and repeated mistakes over time is the next piece to arrive.
Long-only US-listed stocks and ETFs with day limit orders in regular hours: no shorting, no options, no crypto, no margin, and gross exposure capped below your equity. During the beta it works with a US broker that offers a paper-trading API; you bring your own account and connect it yourself after signing in. The trading key is stored sealed and Bellwether never asks for anything that can move money.
You, and anyone you explicitly invite as a read-only guest. Books are isolated from each other at the database layer; the house operator administers logins, not your positions.
Join the waitlist. We are letting people in slowly while the desk runs on paper; when a seat opens you get a one-time setup link by email, choose a password and enrol an authenticator app. There is no app-store download and nothing to install.
Today it runs on paper money in the open, and we are letting people in a few at a time. Leave your email and we will write when a seat opens.
Built by engineers who work on frontier AI systems for a living.