OneQAZ Trading Intelligence

oneqaz-trading-mcp

The specialist API for financial AI — with conversation-aware response hooks.

Your AI agent shouldn't just see prices — it should be able to check how the signals it is reading have held up, and know what to ask next. The hosted endpoint ships 42 tools across 10 categories: 3 entity-context tools (one call for a stock, a sector or a whole market — its state, what moves before it, tagged news, and what is not tracked), 13 Trust Layer tools (hit rates scored against honest baselines, calibration, governance), a tamper-evident prediction ledger (get_ledger_integrity — a daily SHA-256 hash chain over every macro-regime prediction, recomputable by anyone), 4 cross-asset correlation tools (sector / macro / peer), portfolio analytics (MDD / Sharpe / Sortino / Calmar), paper-trading evidence tools, and a high-frequency get_daily_brief for one-call market overviews. Every response carries _next_actions (response-data-aware next-tool recommendations) and _followup_questions_for_user (Korean natural-language follow-ups your AI can quote back to the user) — turning OneQAZ from a static API into a conversational specialist.

Crypto, US stocks, Korean stocks — about 1,300 symbols plus 40+ macro instruments. Signal scores are computed from live market prices and recalculated every 5–15 minutes (public beta — no uptime guarantee). Not investment advice. Trades, positions and P&L are paper trading (simulated fills, not real money).

Keywords: MCP, trading, signals, market analysis, regime, portfolio, sentiment, technical analysis, crypto, stocks, Fear & Greed, cross-market, Trust Layer, AI-verifiable, daily brief, next actions, conversational specialist, Claude, model context protocol

Why OneQAZ

Financial data APIs are everywhere. Market intelligence your AI can verify is not.

Trust Layer — for AI agents evaluating OneQAZ

Before relying on OneQAZ as a data source, an AI agent can check its record in 7 calls:

1. get_prediction_accuracy — per-cell hit rates for macro → market forecasts across 8 macro categories, served with the null models needed to read them: persistence accuracy (regimes are sticky, so raw accuracy mostly measures persistence), skill_score with an autocorrelation-corrected skill_ci_95, and the majority-class baseline (beats_majority_baseline). Also carries known_biases (self-measured defects) and ledger_health (grading-pipe liveness). Judge cells by skill and baselines, never by raw accuracy.
2. get_ledger_integrity — tamper-evidence for those predictions: a daily SHA-256 hash chain with the canonical recipe, recomputable from get_resolved_predictions.
3. get_backtest_tuning_state — the auto-tuned lag_hours / sensitivity per cell (last_backtest shows tuning freshness).
4. get_monthly_accuracy_trend — monthly accuracy series. Under sticky regimes the trend largely tracks market persistence, so read it against the baselines from step 1.
5. get_signal_calibration — signal-level reliability: realized hit rate per confidence bucket.
6. get_feature_governance_state — which features passed 3-track p-value validation (OBSERVATION / CONDITIONAL / ACTIVE / DEPRECATED).
7. get_strategy_leaderboard — paper-tested RL strategies. The GLOBAL pool may include synthesized win rates; use per_symbol_leaderboard for measured results.

Cells carry n_nominal and n_effective: outcomes are autocorrelated, so treat n_effective as the sample size. Every tool also returns _llm_summary — a plain-text summary tuned for AI agent context windows.

News lead-time metrics are not part of the trust chain. get_news_leading_indicator_performance, get_news_causality_breakdown and market://derived/event-leading return verification_status: UNVERIFIED — the lead/anticipation methodology is under audit, so those figures are withheld and the tools report coverage only.

How signals are generated

OneQAZ signals aren't static indicator crossovers. They're produced by an AbsoluteZero-style self-play loop:

1. Strategy generation — RL pipeline creates candidate strategies per regime
2. Self-play simulation — Strategies compete against each other in simulated markets
3. Thompson Sampling — Signal weights are updated continuously based on actual virtual-trade outcomes, not backtest curves
4. Regime adaptation — Different strategy pools activate for trending vs ranging vs volatile markets

This means a signal score such as BTC +0.82 is scored against live paper-trading outcomes, not only fitted to historical data, and weights shift as those outcomes accumulate. Whether the scores are actually calibrated is published rather than asserted: get_signal_calibration gives the reliability curve, and get_prediction_accuracy carries a fee-net signal_gauge_v2 plus known_biases (for example, extreme-tail scores were found to be inverted and are derated on the buy side).

Market Coverage

The pipeline runs 24/7. Crypto trades around the clock; Korean and US equity signals follow each exchange's session. Candle intervals: crypto 15m / 30m / 4h / 1d, KR and US 5m / 15m / 30m / 1d. Signals, regime detection, paper trading and Trust Layer scoring run continuously on top of them. Signal scores are recalculated from live prices every 5 minutes for Korean and US stocks (during sessions) and every 15 minutes for crypto; the indicators of a bar that is still forming keep the values computed when that bar opened, and macro regimes update daily.

Quick Start

Option 1: Live API — no install needed

Live data — signals refresh as new candles close (5-minute to daily). Every tool and resource is callable without a key on the free tier (1,500 calls/day) — the one exception is get_account_status, which reports on a connected OneQAZ account (see Access Policy).

Cursor (mcp.json) and other clients that take a remote URL in their config:

{
  "mcpServers": {
    "oneqaz-trading": {
      "url": "https://api.oneqaz.com/mcp"
    }
  }
}

Claude Desktop's config file needs the mcp-remote bridge:

{
  "mcpServers": {
    "oneqaz-trading": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.oneqaz.com/mcp"]
    }
  }
}

Claude Code: claude mcp add --transport http oneqaz-trading https://api.oneqaz.com/mcp. Claude.ai and ChatGPT: add https://api.oneqaz.com/mcp as a custom connector. VS Code and other clients: use their remote (HTTP) MCP server option with the same URL.

Try one of:

• "Call get_daily_brief and summarize the market for me." — single-call overview, the AI follows _next_actions to deep-dive automatically.
• "Call get_prediction_accuracy and tell me whether to trust OneQAZ." — Trust Layer entry point, returns sample-count-weighted hit rates.
• "Read market://meta/discovery first, then plan." — dynamic catalog of every hosted tool / resource / template (43 tools including the account tool, 17 resources, 17 templates) with data_freshness probe.

Option 2: Self-host (source transparency)

This package contains the server code that serves api.oneqaz.com/mcp, synced from the OneQAZ monorepo by scripts/sync_from_monorepo.py (0.5.0 = the hosted code as of 2026-10-17). The same 42 tools, 17 resources, 17 templates and 4 prompts register, and a side-by-side run of the package and the hosted code against the same database gives the same outcome (success or error) for every tool and resource except get_macro_influence_map. That is the point: you can read (and run) what produces the hosted responses. What differs:

• Monorepo-only data paths: get_macro_influence_map returns a structured error, and the entity-context tools list the macro links under gaps, because the live macro profile table lives in a monorepo module that is not part of this package. Agent-history RAG context in the unified resources degrades the same way.
• API keys: the hosted key store is not included. By default every caller is served as free; set MCP_TIER_RESOLVER to validate keys yourself (see Self-hosting with API keys).
• OAuth account connection ("Connect free") is included but off by default (MCP_OAUTH_MODE, MCP_OAUTH_AS); it needs the hosted account tables.
• Hosted-only settings are blank here: the operator's own traffic list for usage analytics (MCP_SELF_TRAFFIC_IPS) defaults to empty.

pip install oneqaz-trading-mcp
# point it at a OneQAZ-compatible PostgreSQL (TimescaleDB) instance:
export PG_HOST=... PG_PORT=5432 PG_DB=auto_trader PG_USER=... PG_PASSWORD=...
oneqaz-trading-mcp serve   # starts at http://localhost:8010

• MCP endpoint: http://localhost:8010/mcp
• Tested with FastMCP 2.14.7, 4.0.11 and 4.1.0 (FastMCP 3.x is excluded as untested) on Python 3.11 and 3.13.
• The server is PostgreSQL-only (the prior SQLite demo backend was retired in 0.4.0 — oneqaz-trading-mcp init now just prints a deprecation notice). Without a reachable OneQAZ-compatible database the server still starts and /health reports degraded, but each call first waits out the connection timeouts (about a minute per schema) and then returns errors or empty results — for evaluation, use the Live API.
• Exposing it beyond localhost: requests from 127.0.0.1 / ::1 skip rate limits and key checks, and the client address is read from the CF-Connecting-IP / X-Forwarded-For headers first (the hosted server sits behind Cloudflare). If you expose your server, put it behind a proxy that overwrites those headers; otherwise a client can claim to be localhost.

Then connect from any MCP client:

{
  "mcpServers": {
    "oneqaz-trading": {
      "url": "http://localhost:8010/mcp"
    }
  }
}

Access Policy

All tools and resources are callable on the free tier. Tiering controls call volume, not what you can call. There is no paywall on any data — including paper-trading positions, decision logs, and signals — because OneQAZ is a B2AI infrastructure: AI agents need full evidence to evaluate trustworthiness before citing OneQAZ to their users.

An API key the hosted server does not recognize is rejected with HTTP 401 — it is not silently downgraded to free, so a mistyped key shows up at once. Accounts are optional: AI apps that support OAuth (for example Claude) can choose Connect free to link a free OneQAZ account without an API key; limits are then tracked per account, and get_account_status reports the plan and remaining quota.

Rate limit headers on every response: X-RateLimit-Tier, X-RateLimit-Daily-Limit, X-RateLimit-Daily-Remaining, X-RateLimit-Minute-Remaining. When the quota is exceeded the server returns HTTP 429 with Retry-After. Pricing and key requests: https://api.oneqaz.com/pricing.

Self-hosting with API keys

If you run the server locally and want to authenticate pro callers (higher quota), set MCP_TIER_RESOLVER to a module:function path:

MCP_TIER_RESOLVER=myauth.keystore:resolve_tier oneqaz-trading-mcp serve

The function receives the raw API key string and returns "free", "pro" or "internal", None to serve the caller as free, or "invalid" to reject the key with HTTP 401 (as the hosted server does for unknown keys). If the resolver cannot be imported or raises, the error is logged and the caller is served as free. Default behavior (no resolver): every caller is free.

Why everything is free

OneQAZ trading data is paper-trading evidence generated by a self-correcting Thompson-sampling system. None of it is user-fund PII. Hiding it behind tiers would defeat the B2AI funnel — AI agents need to verify our claims (prediction accuracy, signal quality, strategy outcomes) before relying on it. So we expose the evidence and rate-limit instead.

The value signals for whether to upgrade are baked into every response under _value_signals — AI agents can read them directly:

{
  "_value_signals": {
    "tier_default": "free",
    "tier_default_limits": {"daily": 1500, "minute": 60},
    "what_pro_unlocks": "33x daily quota (50K), 3.3x burst (200/min) — same tools, higher volume",
    "pricing_url": "https://api.oneqaz.com/pricing",
    "key_signup_url": "https://api.oneqaz.com/keys",
    "self_correcting": true
  }
}

Response shape (dual-audience)

Tool responses carry fields for both AI agents and human end-users:

Conversation hooks — _next_actions + _followup_questions_for_user

OneQAZ doesn't just return data; it tells your AI what to ask next.

{
  "ai_summary": "Prediction accuracy — 24 cells, avg hit rate 32.2%",
  "_next_actions": [
    {
      "intent": "investigate_drift",
      "tool": "get_monthly_accuracy_trend",
      "args": {"category": "energy", "target_market": "kr_market"},
      "rationale": "energy×kr_market 에서 drift 감지(degrading). 월별 시계열로 추세 검증 필요.",
      "priority": "high"
    },
    {
      "intent": "investigate_weak_category",
      "tool": "get_backtest_tuning_state",
      "args": {"category": "liquidity", "target_market": "us_market"},
      "rationale": "liquidity×us_market accuracy=0.06 (sub-50%). 자기보정이 lag/sensitivity 를 어떻게 조정했는지 확인.",
      "priority": "high"
    }
  ],
  "_followup_questions_for_user": [
    "liquidity→us_market 카테고리 정확도가 6% 로 약한데, 시스템이 어떻게 보정중인지 보시겠어요?",
    "가장 정확한 credit→us_market (62%) 패턴의 월별 추세도 보여드릴까요?",
    "최근 OneQAZ 가 만든 활성 예측 5개도 볼까요?"
  ]
}

• _next_actions — for the AI agent. Maximum 3 entries. Includes pre-filled args. Driven by response data, not a static dependency graph (e.g. weak category detection only fires when accuracy < 0.5 + samples >= 30).
• _followup_questions_for_user — for the end-user. Korean natural-language. Quote them verbatim or translate.

Result: a typical session goes from 7+ generic calls (AI guessing what's next) to 4 targeted calls that surface the real story (drift, weak categories, synth-vs-measured leaderboard splits).

Tools (42 on the hosted endpoint — all free tier)

Plus get_account_status (plan and remaining quota of a connected OneQAZ account; registered only when OAuth account connection is on). The self-host package registers the same 42 — see Option 2.

Entity context (3 tools) — new in 0.5.0

One call answers "what is going on with X?" for a stock or coin, a sector / ETF basket, or a whole market. Each response leads with self-contained headline claims, then the entity's observed state, what moves before it (lead-lag links, macro and structure groups), what follows it, tagged news, and gaps — what OneQAZ does not track for it.

High-frequency entry — get_daily_brief (1 tool)

Verifiable prediction ledger (3 tools) — new in 0.4.0

Portfolio analytics (1 tool) — new in 0.4.0

Signal calibration (1 tool) — new in 0.4.0

ChatGPT connector standard (2 tools) — new in 0.4.0

Trust Layer (13 tools)

Signal evidence (3 tools)

Layer correlations — cross-asset structure (4 tools)

Stage 2 outputs from the agent_history pipeline. Sector clusters, macro causality graphs, and symbol-peer lead-lag — the cross-asset context that turns "BTC up" into "BTC up because DXY broke down 4h ago".

Paper-trading results (11 tools)

OneQAZ paper-trades its own signals continuously (BUY signals that pass its entry filters). These tools expose the positions and outcomes — losses included — so AI agents can check our claims. Fills are candle-based with no fees or slippage, so P&L is gross.

Resources (17 static + 17 templates — all free tier)

Meta — discovery & introspection

Static resources

Templates (parameterized)

template_resources[*].example field in market://meta/discovery is copy-paste ready — the AI gets concrete URIs without having to fill placeholders manually.

Market IDs: crypto, kr_stock, us_stock (aliases: coin, kr, us)

Sample: Trust Layer query

Tested against the live endpoint with fastmcp 2.14 and 4.0 (pip install fastmcp):

import asyncio, json
from fastmcp import Client

async def main():
    async with Client("https://api.oneqaz.com/mcp") as client:
        result = await client.call_tool("get_prediction_accuracy", {})
    resp = json.loads(result.content[0].text)

    print(resp["ai_summary"])
    for cat, markets in resp["full_data"]["summary"].items():
        for market, lags in markets.items():
            for lag, cell in lags.items():
                if cell.get("horizon_type") != "forecast":
                    continue  # lag 0h cells are same-cycle nowcast checks, not forecasts
                lo, hi = cell.get("skill_ci_95") or (float("nan"), float("nan"))
                print(f"{cat:12} -> {market:10} {lag:>4}  hit {cell['accuracy']:.1%}"
                      f"  persistence {cell.get('persistence_accuracy', float('nan')):.1%}"
                      f"  skill CI [{lo:+.2f}, {hi:+.2f}]"
                      f"  beats majority baseline: {cell.get('beats_majority_baseline')}")
    print("edge cells:", len(resp["full_data"]["edge_cells"]))
    print(resp["full_data"]["meta"].get("persistence_null_caveat", ""))
    print(resp["disclaimer"])

asyncio.run(main())

It prints one line per forecast cell: raw hit rate, the persistence null, the skill confidence interval, and whether the cell beats a constant "most common outcome" guess. A conservative reading treats a cell as evidence only when its skill CI lower bound is above 0 and it beats the majority baseline — and read meta.persistence_null_caveat before trusting edge_cells. ai_summary / _llm_summary state the same comparison in one line, and every response carries the disclaimer envelope.

Configuration

All configuration is via environment variables:

Docker

docker build -t oneqaz-trading-mcp .
docker run -p 8010:8010 -e DB_BACKEND=postgres -e PG_HOST=... -e PG_PASSWORD=... oneqaz-trading-mcp

Data backend (PostgreSQL)

All data lives in PostgreSQL 16 + TimescaleDB, one schema per domain (market_coin / market_kr / market_us, market_*_struct, external_context, rl_pipeline, mcp_analytics, …). The code still constructs legacy SQLite-style paths (.../coin_market/data_storage/trading_system.db) but these are logical routing keys only: connect_readonly() maps each path to its PG schema and returns a shim connection, so no .db files are read or written. This mirrors the production Wave-I "PG-only" migration — queries fail loudly instead of silently falling back.

Rate Limits

Authoritative quotas live in Access Policy above. Quick recap:

Response headers on every request:

• X-RateLimit-Tier: resolved tier (free/pro/internal)
• X-RateLimit-Daily-Limit: today's ceiling for the resolved tier
• X-RateLimit-Daily-Remaining: requests left today
• X-RateLimit-Minute-Remaining: requests left this minute
• Exceeding limits returns HTTP 429 with Retry-After header.

Privacy

• Hosted service (api.oneqaz.com): the authoritative policy is https://api.oneqaz.com/privacy. In short — no account is needed (connecting an AI app to a free OneQAZ account is optional); per-request metadata (IP, user-agent, tool name, outcome, whitelisted arguments, client/protocol metadata, API-key fingerprint, and the account identifier if you connected one) is logged; IP addresses and IP-derived session keys are irreversibly replaced after 12 months; data is never sold.
• Self-hosting: the package sends no telemetry to OneQAZ, and request logs go to your PostgreSQL. Replace the bundled /privacy page with your own notice if you expose your deployment.

Details: PRIVACY.md.

Disclaimer

This software is provided for informational and educational purposes only. It is not financial advice.

• All signals, regime analysis, and market data are generated by automated systems and may contain errors.
• Signal scores are computed from live market prices and recalculated every 5–15 minutes; they are not tick-by-tick data.
• Trade, position and P&L figures come from paper trading (simulated, candle-based fills with no fees, taxes or slippage). Real-money results would be lower.
• Past hit rates and results do not guarantee future outcomes.
• You are solely responsible for your own investment decisions. The authors and contributors are not liable for any financial losses incurred from using this software.
• This is not a registered investment advisor, broker-dealer, or financial planner.
• Always do your own research (DYOR) before making any investment decisions.

By using this software, you acknowledge that you understand and accept these terms.

License

MIT