What's new

Claude Opus 5 — and a model picker per agent

Opus 5, Anthropic's newest frontier model, is now the default for new agents — at the same per-token price as the 4.8 it replaces. Which model your agent thinks with is your setting, not our decision.

• Choose per agent between Opus 5, Opus 4.8, Claude Fable 5 and the GPT-5.6 line — each shown with its typical cost per run before you commit to it.
• The trader, the controller and the curator can each run a different model: pair a fast, cheap trader with a stronger controller that holds veto power over it.
• Agents already running keep the model they were built on until you change it.

Every closed trade, on the chart

Trade history stopped being a table of numbers. Expand any closed trade to see the candles around it — entry, exit and the levels the agent was working with, on the timeframe it was actually watching.

• Average hold time now scales to the trade: minutes, hours or days, whichever reads honestly.
• Every run carries the market context it saw and the timestamp it saw it at — so a decision can be audited against the tape instead of from memory.

A paper balance you can reconcile by hand

Paper accounts now compute balance as one line of arithmetic: your starting balance plus the summed realised PnL of your closed trades. Nothing else moves it.

• No simulated fees, no slippage, no funding, and no margin locked out of the balance while a position is open — the number on your dashboard and the numbers in your trade list can no longer disagree.
• Existing paper accounts were recomputed onto the new basis.
• The trade-off is deliberate: paper now measures your strategy's decisions cleanly, rather than approximating an exchange's cost structure on top of them.

Aggressiveness rewrites the strategy, not a number

Aggressiveness is no longer a dial the agent reads and interprets — it re-renders the doctrine itself. Conservative and balanced keep the full guard gates. Aggressive hands the timing calls back to the trader. Max drops the setup checklist entirely and makes discretion the strategy.

• Looser tiers remove trend-maturity caps, allow chasing with a structural stop, and permit re-entry on fresh evidence — the things a strict template forbids by construction.
• Your explicit constraints — long-only, short-only, no counter-trend — survive every tier. Those are your rules, not guard gates.
• Still separate from sizing: how loosely an agent enters and how much it risks stay two different settings.

The strategy wizard — and a controller with veto power

A second way in: describe how you trade in plain language and a system-aware interviewer turns it into doctrine — telling you honestly when the mechanic you're describing isn't one the engine supports. It writes the strategy only once you confirm it.

• Execution became two knobs per agent: let deterministic code size the position from your risk settings, or let the controller reason about size itself inside a written policy.
• Optionally the controller becomes a bad-idea filter — it reads your strategy and the chart and can veto a trade the trader wanted. A veto wakes the trader to reassess rather than silently dropping the idea.
• The controller now runs as its own metered step alongside the trader, so a heavier controller doesn't slow a fast trader down.

Paper mode, everywhere, by default

Every new system now starts on a simulated per-system account that trades against real market data — full order, position, TP/SL and liquidation mechanics, none of your capital. Going live is a separate, confirmed switch that requires a connected exchange key and a flat book.

• Simulated fills at mark ± slippage with taker fees and approximate isolated-margin liquidation.
• A background settler evaluates your stops and targets against live candles even while the agent sleeps — an exchange executes your stop whether or not you're awake.
• Reset a paper account to a clean balance anytime before you go live.

The Learning Curator

A second, slow-cadence agent that never trades. On its own schedule it reviews the period's trades and market behavior against your strategy and — optionally — curates a single, versioned playbook the trader reads on every run.

• Your ETH agent and your SOL agent now genuinely diverge over time — each learns how its own asset moves under your doctrine.
• The learning loop is opt-in per system; turn it off and the trader runs on your doctrine alone.
• A "lesson" that tries to relax a hard invariant is flagged, never silently followed.

The price-trigger watchtower

Agents no longer wake only on a timer. Each run can arm price levels it cares about — and the moment price crosses one, the agent wakes to ask its one narrow question. Patience without blindness.

• Watching a setup now means "wake me at this level," not "re-scan the whole chart every hour."
• A moved stop can't fire off pre-move wicks — triggers only evaluate candles since the level was armed.

Prepaid compute wallet + auto-reload

Billing landed: a monthly platform seat plus a prepaid wallet metered at the token level per run. You pay for what your agents actually use — and patient agents run rarely.

• Stripe-hosted checkout and billing portal — no card details ever touch the app.
• Optional auto-reload with per-day caps and a trip switch after repeated declines.
• Low balance? The agent pauses. It never trades broke.

Promoter / Controller split

Judgment and execution now live in separate hands. The agent that decides whether to trade emits a signal only — side, entry, stop, target, confidence — and can't size or place a single order. A separate controller owns risk, with hard gates the agent never sees.

• Position sizing, leverage caps and risk gates are deterministic code, not an LLM guess.
• A trade that fails a risk gate is rejected before any model even runs.

Bring-your-own-keys onboarding

Connect your own exchange account with trade-only API keys, encrypted at rest in a managed vault and decrypted only at run time. Start from a strategy template or write your own doctrine in plain language — a guided interview checks it for consistency before it can go live.

Closed beta opens

The first founders get their seats. A single autonomous agent, a state machine of Idle → Watching → In-trade, waking on a schedule to ask one question at a time — the discipline you lose in the market, running while you sleep.