Multi-agent market simulation framework for traders. Built on Mastra so you can pick any LLM provider Mastra supports (OpenAI, Anthropic, Google, DeepSeek, Groq, Mistral, xAI, OpenRouter, Ollama, etc.) and gain Memory / Workflows / Scorers / Observability for free. Persistence via bun:sqlite.
⚠️ Bun-only. Sealion usesbun:sqlitedirectly. Node.js is not supported — install Bun ≥ 1.1.
bun add sealion-simimport {
ActionType,
TraderGraph,
LLMAction,
TradingAgent,
TraderProfile,
make,
spotPreset,
} from "sealion-sim";
const traders = new TraderGraph();
traders.addAgent(
new TradingAgent({
agentId: 0,
profile: new TraderProfile({ handle: "alice", initialCash: 10_000 }),
model: "openai/gpt-4o-mini", // or anthropic/, google/, groq/, deepseek/...
}),
);
const env = make({
traders,
market: spotPreset({ symbol: "BTC/USDT" }),
databasePath: "./data/sim.db",
});
await env.reset();
await env.step(new Map([[traders.getAgent(0), new LLMAction()]]));
await env.close(); ┌──────────────────────────────────────┐
│ Scenario │
│ market + pricer + traders + policy │
└───────────────┬──────────────────────┘
│ .run()
▼
┌──────────────────────────────────────────────────────┐
│ SealionEnv │
│ reset / step / close lifecycle │
└───────┬───────────────────────────────────┬──────────┘
│ tools dispatch │ pricer.next() each step
▼ ▼
┌────────────────────┐ ┌────────────────────┐
│ TradingAgent │ ──Channel──▶ │ Exchange │
│ (Mastra Agent) │ ◀─────────── │ + MatchingEngine │
│ + MarketMakerBot │ │ + Risk gate │
└────────────────────┘ │ + SQLite │
└────────────────────┘
TraderGraph API:
graph.addAgent(agent);
graph.removeAgent(agent);
graph.getAgent(agentId);
graph.getAgents(); // → Array<[id, agent]>
graph.getAgents([0, 1, 2]); // filter by ids
graph.hasAgent(agentId);
graph.getNumNodes();
// copy-trade edges (optional):
graph.addEdge(leaderId, followerId, ratio);
graph.removeEdge(followerId);
graph.getEdges(); // → [leaderId, followerId, ratio]
graph.getNumEdges();
graph.getFollowers(leaderId);
graph.getLeader(followerId);
graph.isLeader(agentId);
graph.isCopying(agentId);src/
├── environment/
│ ├── env.ts SealionEnv (reset/step/close)
│ ├── env-action.ts ManualAction / LLMAction
│ └── make.ts make() factory
├── trading-agent/
│ ├── agent.ts TradingAgent (wraps Mastra Agent)
│ ├── agent-action.ts Mastra createTool() per action
│ ├── agent-environment.ts Per-agent prompt/data view
│ ├── agent-graph.ts TraderGraph (nodes + copy-trade edges)
│ ├── agents-generator.ts ARCHETYPES + factories
│ └── market-maker-bot.ts Rule-based auto-quoting bot
├── platform/
│ ├── platform.ts Exchange (matching + dispatch)
│ ├── channel.ts Actor mailbox (FIFO per agent)
│ ├── database.ts SQLite schema + helpers
│ ├── matching-engine.ts Price-time priority + self-trade prevention
│ ├── pricer.ts 8 price-process simulators
│ ├── typing.ts ActionType / OrderSide / etc.
│ └── config/trader-info.ts TraderProfile + persona blocks
├── clock/clock.ts SandboxClock
├── testing/show-db.ts printDbContents
├── scenarios/scenario.ts High-level orchestrator
├── markets/presets.ts spot / perp / equity / custom
├── evals/scorers.ts Mastra scorer factories (PnL, persona)
├── mastra-integration.ts collectMastraAgents, observability config
└── metrics.ts Sharpe, max DD, FIFO win rate
Anything Mastra's model router supports — string-shorthand "provider/model":
new TradingAgent({ ..., model: "openai/gpt-4o-mini" });
new TradingAgent({ ..., model: "anthropic/claude-sonnet-4.5" });
new TradingAgent({ ..., model: "google/gemini-2.0-flash" });
new TradingAgent({ ..., model: "groq/llama-3.3-70b-versatile" });
new TradingAgent({ ..., model: "deepseek/deepseek-chat" });
new TradingAgent({ ..., model: "openrouter/meta-llama/llama-3.1-405b" });
new TradingAgent({ ..., model: "ollama/llama3.1" });
// ...100+ providers via MastraOr pass an AI SDK LanguageModel instance, or any Mastra MastraModelConfig.
import { Memory } from "@mastra/memory";
new TradingAgent({
agentId: 0,
profile,
model: "openai/gpt-4o-mini",
memory: new Memory({ options: { lastMessages: 20, observationalMemory: true } }),
});import { Mastra } from "@mastra/core";
import { LibSQLStore } from "@mastra/libsql";
import { Observability } from "@mastra/observability";
import { collectMastraAgents, sealionObservabilityConfig } from "sealion-sim";
export const mastra = new Mastra({
agents: collectMastraAgents(traders),
storage: new LibSQLStore({ url: "file:./mastra.db" }),
observability: new Observability(sealionObservabilityConfig({})),
});Studio menampilkan setiap trade decision, tool call, token usage, dan trace.
import { createPnLScorer, createPersonaConsistencyScorer } from "sealion-sim";
const pnl = createPnLScorer({ baselineEquity: 10_000, targetEquity: 15_000 });
const result = await pnl.run({
db: env.exchange.db,
agentId: 0,
startingCash: 10_000,
});
// → { score: 0.42, reason: "...", metadata: { totalReturnPct, sharpe, maxDrawdownPct, ... } }Bisa juga di-attach langsung ke TradingAgent:
new TradingAgent({
...,
scorers: {
pnl: { scorer: pnl, sampling: { type: "ratio", rate: 1 } },
},
});ActionType.PLACE_LIMIT_ORDER; // { side, price, size }
ActionType.PLACE_MARKET_ORDER; // { side, size }
ActionType.CANCEL_ORDER; // { order_id }
ActionType.CANCEL_ALL_ORDERS; // {}
ActionType.GET_ORDERBOOK; // { depth? }
ActionType.GET_RECENT_TRADES; // { limit? }
ActionType.GET_ACCOUNT; // {} → cash, position, avg, P&L, equity
ActionType.GET_OPEN_ORDERS; // {}
ActionType.DO_NOTHING; // {}
ActionType.COPY_TRADE; // { leader_id, ratio }
ActionType.UNFOLLOW; // {}
ActionType.SHARE_SIGNAL; // { content, symbol?, bias? }
ActionType.GET_SIGNALS; // { limit? }
ActionType.INTERVIEW; // { prompt, response }Default subsets: DEFAULT_TRADER_ACTIONS, SOCIAL_TRADER_ACTIONS, SOCIAL_PLUS_TRADING_ACTIONS.
All deterministic given a seed.
new ConstantPrice({ price: 100, steps: 50 });
new RandomWalk({ start: 100, sigma: 1, drift: 0.01, seed: 42 });
new GBM({ start: 50_000, mu: 0.0001, sigma: 0.02, seed: 42 });
new MeanReversion({ start: 100, mu: 100, theta: 0.1, sigma: 1, seed: 42 });
new Replay({ ticks: [{ price: 100 }, { price: 101 }, ...] });
new JumpDiffusion({ base: gbm, jumpProb: 0.05, jumpStd: 0.1 });
new Composite([calmRegime, volatileRegime]);
new Scripted({ base, shocks: [{ atStep: 50, multiplier: 0.85 }] });spotPreset({ symbol: "BTC/USDT" });
perpPreset({ symbol: "BTC-PERP" });
equityPreset({ symbol: "AAPL" });
customMarket({ symbol, baseAsset, quoteAsset, tickSize, lotSize, takerFeeBps, makerFeeBps });import { Scenario, GBM, Scripted, spotPreset, makeMarketMaker } from "sealion-sim";
const traders = new TraderGraph();
traders.addAgent(makeMarketMaker({ agentId: 0, cash: 1_000_000, position: 100 }));
traders.addAgent(new TradingAgent({ ... }));
const sc = new Scenario({
market: spotPreset({ symbol: "BTC/USDT" }),
databasePath: "./data/sim.db",
fresh: true,
traders,
pricer: new Scripted({
base: new GBM({ start: 50_000, mu: 0, sigma: 0.02, seed: 1 }),
shocks: [{ atStep: 50, multiplier: 0.85 }], // -15% flash crash
}),
steps: 100,
decideAction: (agent, ctx) => myStrategy(agent, ctx),
});
const result = await sc.run();
// → { steps, metrics: Map<id, Metrics>, accounts, env }bun run examples/spot-manual.ts # manual orders, no LLM
bun run examples/position-sim.ts # hold a position, watch P&L vs GBM
bun run examples/scenario-flash-crash.ts # scripted -15% shock
bun run examples/copy-trade.ts # leader fill → follower mirror
bun run examples/backtest.ts # Replay synthetic OHLCV
OPENAI_API_KEY=sk-... bun run examples/spot-llm.tsLihat examples/README.md untuk overview tiap demo.
bun run lint # oxlint
bun run lint:fix
bun run format # oxfmt --write
bun run format:check
bun run typecheck # tsc --noEmit
bun run check # format + lint + typecheck
bun test # 26 tests covering engine, copy-trade, pricers, risk, scenarios- No global defaults —
MarketSpecselalu eksplisit lewat preset. - Lazy LLM agent — Mastra
Agentcuma diinstansiasi saat aksi LLM dipakai, jadi sim pure-manual tidak butuh API key. - Single-writer SQLite — semua mutasi diserialisasikan oleh
Exchange.runLoop()lewatChannel. - Self-trade prevention — matching engine otomatis exclude order dari agent yang sama.
- Cached mark price — invalidasi otomatis pada injectPrice / fill.
- Skip read-only trace —
GET_*actions tidak nge-spam trace table. - FIFO win-rate —
computeMetricspair fills properly bukan running-avg approximation. - Apa pun bisa disimulasikan — harga, posisi, regime shift, copy-trade, news shock, multi-LLM benchmark — semua dari building blocks yang sama.
Apache-2.0.