OpenAlphaStack is an open-source Codex stack for A-share research. Domain Skills encode reusable methods, MCP exposes typed tools, and a deterministic Python engine owns T+1 rules, fees, risk, state, and paper execution.
{
"ok": true,
"data": {
"research_mode": "paper",
"market_bias": "neutral"
},
"meta": {
"source": "provider-or-demo",
"as_of": "2026-07-24T10:00:00+08:00",
"freshness": { "status": "fresh" },
"demo": false
}
}Not an AI stock oracle
One model should not impersonate an analyst, risk engine, exchange, and auditor at the same time. OpenAlphaStack gives each layer a narrow job and makes missing, stale, or failed data visible.
Market regimes, sector rotation, candidate screening, and stock review become reusable workflows instead of prompts trapped in chat history.
One Codex task composes Skills by default. Multi-agent orchestration is optional, not theater.
Quotes, technicals, fundamentals, news, risk calculations, and baseline backtests use typed inputs and versioned responses.
Every market result preserves source, as-of time, freshness, and synthetic-demo status.
T+1 behavior, fees, position limits, idempotency, state transitions, and ledger commits are deterministic code paths.
Public modes cover research, backtesting, and paper trading. There is no live broker integration.
Four explicit layers
The agent proposes and explains. Tools retrieve. Rules execute. The ledger preserves evidence.
Interpret research intent, compose methods, cite evidence, and produce reviewable conclusions.
Encode market analysis, screening, stock research, and intraday guardrails.
Expose market data, indicators, news, backtests, and bounded paper-plan operations.
Enforce invariants, commit state atomically, and preserve append-only audit records.
Inspect before trusting
The repository ships a reproducible synthetic demo, a typed MCP contract, a paper-only engine, a local observability dashboard, tests, release notes, and contributor entry points.
Agent confidence = audit metadata
Research narrative = reviewable context
Execution gates:
✓ validated plan schema
✓ deterministic risk rules
✓ idempotency key
✓ expected state version
Never exposed:
× live broker orders
× arbitrary shell access
× fabricated missing dataReach first success quickly
You do not need a market-data provider or a running paper engine to verify installation, Skill discovery, and MCP calls. Start with deterministic synthetic data, then move to live research.
git clone https://github.com/44-99/OpenAlphaStack.git
cd OpenAlphaStack
py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .
openalphastack doctorUse the market-analysis Skill in offline demo mode. Read the market overview and market news through OpenAlphaStack MCP, and clearly label every value as synthetic demo data.
Boundaries over slogans
Backtests and paper results are research evidence. OpenAlphaStack helps you inspect assumptions, data, rules, state, and outcomes; it does not turn uncertainty into a guaranteed answer.
Frequently asked questions
No. Public modes cover research, backtesting, and paper trading. No live broker tool is exposed.
One Codex task can compose the domain Skills for most research. Parallel agents are reserved for explicitly independent work.
No. This site is static, and the local dashboard binds to 127.0.0.1 by default. The separate public MCP surface is stateless and read-only.
Yes. The repository includes deterministic synthetic datasets for installation and workflow verification. They are never presented as investment data.
Start with reproducibility
Inspect the source, complete the offline demo, and verify the boundaries. If the architecture is useful, star the repository, report a problem, or contribute a focused improvement.