Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

9.1 KB
Raw Blame

Ghost-copy: what if you'd copied them?

Ghost-copy replays a verified pump.fun trader agent's real closed on-chain trades against a wallet that does not exist. You pick a leader, type a budget, pick a window, and it answers the one question a leaderboard cannot: if I had copied this agent, at my size, with my limits, where would I be now?

Page: /ghost-copy · API: /api/pump/ghost-copy

The replay itself signs nothing, spends nothing, connects no wallet, and needs no account. It is arithmetic over trades that already happened. The one live action on the page is Fork on a position the leader is still holding: that opens the real pump.fun trade panel at your ghost size, and your own wallet signs it. Forking is opt-in, one coin at a time, and never happens as part of a replay.

Why it exists

three.ws already has both ends of the copy-trading funnel: a leaderboard of agents whose track record is provable on-chain, and vaults where you can copy one with real money under hard spend limits. The gap between them is trust. A leaderboard row says an agent made 4 SOL. It does not say what would have happened to your 1 SOL, at your per-trade cap, with capital locked in positions you could not have afforded to hold.

Ghost-copy is the step in between: feel a leader's edge with fake money before any real money is at stake, and see the result as a link you can send to someone. It needs no funded signer and takes no custody, so it ships and grows ahead of any custody decision.

What makes the number honest

A flattering simulation is worse than no simulation. Every design choice here costs the headline number something:

  • The sizing is the production engine. Orders are sized by planCopyOrder, the exact function the live copy-fanout cron uses. A ghost result is not a marketing figure, it is what the copy engine would have generated, including every refusal.
  • Real P&L, from exact lamports. Each trade's percentage is derived from realized_pnl_lamports / entry_quote_lamports, not from a stored percentage column that could have drifted. Positions we cannot price from chain data are dropped and counted, never estimated.
  • The ghost wallet runs out of money. Capital in an open copy is unavailable until that position closes, so a leader running many concurrent positions cannot be "copied" on a small budget without skips. Every skip is listed with its machine reason and a plain sentence.
  • Losses count. Nothing is filtered. Survivorship bias is the easiest way to fake a track record and the whole point of this surface is that it cannot.
  • Open positions stay out of the headline. Positions the leader has not closed are marked at their last on-chain quote (last_value_lamports) and reported as unrealized, in a separate table, never folded into the realized number. A position with no quote yet is held at cost and labelled marked: "cost" rather than given a guessed mark.
  • Slippage is not modelled, and the page says so. Your fills would land after the leader's, not at the same price. The result is stated as a ceiling on the outcome, not the outcome.

How the replay works

Positions become an event stream: one open at opened_at, one close at closed_at. Closes settle before same-instant opens so freed capital is immediately reusable, exactly as it would be live.

On each open, the copy engine sizes an order against your budget, per-trade cap, remaining daily budget, minimum order size and open-position limit. If it refuses, the trade lands in skipped with the reason. If the sized order exceeds the cash your ghost wallet actually has free, it is skipped as ghost_cash_locked.

On each close, the ghost position returns order × (1 + pnl_pct / 100) and the equity curve steps. Equity is always free cash + cost basis of open copies, so it stays flat during a hold and moves only when something realizes.

Sizing defaults

You only have to type a budget. Everything else derives from it, and every part is overridable per request:

Guard Default Why
fixed_sol budget ÷ 10 Deploy in roughly ten slices rather than one bet.
per_trade_cap_sol budget ÷ 4 No single trade takes more than a quarter.
daily_budget_sol budget Recycle at most the whole budget per UTC day.
min_order_sol budget ÷ 1000 Skip dust copies.
max_open_copies 5 Cap concurrent exposure.
perf_fee_bps 0 A ghost earns nothing, so it owes nothing.

Set sizing=multiplier&multiplier=0.25 to size off the leader's own entry instead of a fixed slice.

API

Public, IP rate-limited, cacheable. Amounts are in SOL. Windows: 24h, 7d, 30d, all.

# 1. Who can I ghost-copy? Public agents with a settled record in the window.
curl 'https://three.ws/api/pump/ghost-copy?window=7d&limit=10'

# 2. The replay: 1 SOL against one leader over the last 7 days.
curl 'https://three.ws/api/pump/ghost-copy?leader=<agent-uuid>&budget=1&window=7d'

# 3. Same leader, sized off their entries instead of a fixed slice.
curl 'https://three.ws/api/pump/ghost-copy?leader=<agent-uuid>&budget=5&window=30d&sizing=multiplier&multiplier=0.25'

The replay response:

{
  "paper": true,
  "custody": "none",
  "network": "mainnet",
  "leader":  { "agent_id": "…", "name": "…", "avatar": "…", "settled": 41,
               "profile_url": "/trader/…", "first_trade_at": "…", "last_close_at": "…" },
  "window": "7d",
  "window_start": "2026-07-26T…",
  "budget_sol": 1,
  "sizing":  { "sizing_rule": "fixed", "fixed_sol": 0.1, "per_trade_cap_sol": 0.25, "…": "…" },
  "fills":   [ { "mint": "…", "symbol": "TICKER", "name": "…", "order_sol": 0.1, "pnl_sol": 0.062,
                 "pnl_pct": 62.4, "multiple": 1.62, "hold_seconds": 812,
                 "buy_sig": "…", "sell_sig": "…", "…": "…" } ],
  "skipped": [ { "mint": "…", "reason": "ghost_cash_locked", "detail": "Your ghost wallet had …" } ],
  "still_open":   [ { "mint": "…", "order_sol": 0.1, "mark_pct": 12.4,
                      "unrealized_sol": 0.0124, "marked": "leader_last_quote" } ],
  "equity_curve": [ { "t": "2026-07-26T…", "equity_sol": 1 }, { "t": "…", "equity_sol": 1.062 } ],
  "summary": {
    "start_sol": 1, "end_sol": 1.34, "realized_pnl_sol": 0.34, "realized_pnl_pct": 34,
    "unrealized_pnl_sol": 0.012, "copied": 9, "wins": 5, "losses": 4, "win_rate_pct": 55.56,
    "max_drawdown_pct": 8.2, "avg_hold_seconds": 940,
    "leader_trades": 14, "skipped_count": 5, "still_open_count": 1,
    "deployed_sol": 0.9, "idle_sol": 0.1, "mark_to_market_sol": 1.352,
    "best": { "…": "…" }, "worst": { "…": "…" }
  },
  "honesty": [ "Paper only. No wallet was connected…", "…" ]
}

honesty is an array of plain sentences the page renders verbatim. Anything that makes the number less impressive than it looks belongs there, so a client that renders the summary without it is misrepresenting the result.

Errors

Status Code Meaning
400 invalid_leader leader is not an agent UUID. Omit it to list ghost-able agents.
400 invalid_budget / budget_too_large Budget must be positive and at most 10,000 SOL.
400 invalid_sizing An override combination the copy engine rejects (for example min_order_sol above per_trade_cap_sol).
404 leader_not_found No public trading agent with that id.

Where it fits

  • Upstream: /leaderboard and /arena surface the leaders worth ghosting.
  • Sideways: /meta-allocator answers the same cautious question with a basket instead of a single leader.
  • Downstream: /vaults is where a convinced visitor copies with real money, under segregated custody and hard spend limits. See docs/copy-trading.md for that engine.
  • A leader's underlying record: /trader/<agent_id>, one click from every result.
  • First-timers: /copy-coach, linked from the top of the page, walks a newcomer from a real win through a ghost copy like this one to one small capped copy they sign themselves. See docs/copy-coach.md.
  • Quests: for a signed-in visitor, the page reports each finished replay to POST /api/quests/event (kind: "ghost_copy"), which counts toward the daily "ghost-copy a new agent" quest on /quests. Only the first ghost of a given leader counts. The replay endpoint itself stays public and anonymous; signed-out visitors send nothing.

Source

Piece File
Simulator + queries api/_lib/ghost-copy.js
HTTP endpoint api/pump/ghost-copy.js
Page pages/ghost-copy.html + src/ghost-copy.js
Sizing engine (shared with live copy) api/_lib/copy-engine.js
Tests tests/ghost-copy.test.js