Demo mode — every number on this page is simulated.The real lab trades a small, ring-fenced experimental balance. This is not a performance claim or investment advice.
← Back to portfolio
Autonomous trading · guardrailed agent

Claude Trading Lab

A spot-only crypto experiment where an AI agent researches, decides and places limit orders on a schedule — inside guardrails a human set once. The interesting part isn't the returns; it's the operating model: journaled runs, hard caps, a kill switch and a weekly review the agent writes about itself.

Market check every 6h3 risk sleevesSpot only · limit ordersWeekly self-review
Steady 25%
Slow allocation into large caps
Rebalances on drift · no leverage · widest patience
Active 50%
Signal-driven swings in large caps
Take-profit + stop on every entry
Spec 25%
Researched entries in high-vol coins
Wider TP/stop · smallest sizing
Equity vs. benchmarks — simulated, 6 weeksindexed to 100
Lab (all sleeves)BTC buy & hold50/50 BTC/ETHCash
Guardrails — set once by the owner
✓
Spot only, limit onlyNo margin, no derivatives, no market orders.
✓
Ring-fenced portfolioTrades a named sub-portfolio; the main account is untouchable.
✓
Max 4 trades / dayExtra only with documented, exceptional evidence.
✓
Sleeve split enforced25 / 50 / 25 — an order that breaks it is rejected before it's placed.
✓
Health before returnsMissed runs are the first line of every report and trigger a push.
✓
Journal everythingEvery run, order, deposit and benchmark row lands in Postgres.
Kill switch — one boolean in lab_settings. When true, the agent reads it first and does nothing but report.
Run loglast 24h
    Order journalsample
      Weekly review — written by the agentSunday 6:00 PM ET

      Health

      27 of 28 expected market-check runs completed this week. One missed run (Wed 01:00 ET) — infrastructure timeout, no orders were pending. Kill switch: off all week.

      Result, simulated

      Combined equity +2.1% for the week vs BTC hold +1.4%, 50/50 +0.9%, cash 0.0%. Steady: +0.8%. Active: +3.6% (two take-profits hit, one stop). Spec: −1.2% (one position stopped out at −6%, within its wider band).

      What I'd change

      Active sleeve entries clustered in the same 6-hour window on three days; I recommend staggering entries across two checks to reduce timing risk. This is a recommendation for the owner — I do not change lab_settings myself.

      Decisions waiting on the owner

      None this week.