# oneqaz-trading-mcp

<!-- mcp-name: io.github.wnsod/oneqaz-trading-mcp -->

[![GitHub stars](https://img.shields.io/github/stars/oneqaz-trading/oneqaz-trading-mcp?style=social)](https://github.com/oneqaz-trading/oneqaz-trading-mcp)
[![PyPI](https://img.shields.io/pypi/v/oneqaz-trading-mcp)](https://pypi.org/project/oneqaz-trading-mcp/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

> **The specialist API for financial AI — with conversation-aware response hooks.**
>
> Your AI agent shouldn't just see prices — it should be able to *prove* the
> signals it's acting on have worked, 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. 24/7 live.
> Research information from paper trading — not investment advice.

**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.

| | Typical financial MCP | OneQAZ |
|---|---|---|
| Price / OHLCV data | ✅ | ✅ |
| Technical indicators | ✅ | ✅ |
| **Regime detection** (trending / ranging / volatile) | ❌ | ✅ |
| **Self-correcting signals** (Thompson Sampling on paper-trade outcomes) | ❌ | ✅ |
| **Macro regime + market-structure overlays** (two independent loops, joined at each symbol's decision) | ❌ | ✅ |
| **Trust Layer** (hit rates vs persistence & majority-class baselines + calibration) | ❌ | ✅ |
| **Self-disclosed defects** (`known_biases`, grading-pipe `ledger_health`) | ❌ | ✅ |
| **Feature governance** (3-track p-value validation) | ❌ | ✅ |
| **Cross-asset correlations** (sector / macro / symbol-peer lead-lag) | ❌ | ✅ |
| **Conversation hooks** (`_next_actions` + `_followup_questions_for_user`) | ❌ | ✅ |
| **Single-call market brief** (`get_daily_brief`) | ❌ | ✅ |
| **Dynamic discovery** (introspection + `data_freshness` probe) | ❌ | ✅ |
| **Hosted 24/7 endpoint** (no API key needed) | ❌ | ✅ |

### Trust Layer — for AI agents evaluating OneQAZ

Before any AI recommends OneQAZ as a data source, it can self-verify 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 such as "BTC BUY 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

| Market | Exchange / source | Universe | Symbols |
|--------|-------------------|----------|---------|
| Crypto | Bithumb | All KRW-listed pairs | ~460 |
| Korean Stocks | KOSPI / KOSDAQ | KOSPI 200 + KOSDAQ 150 | ~350 |
| US Stocks | NYSE / NASDAQ | S&P 500 | ~500 |
| Macro | rates, FX, futures, ETFs | bonds · credit · FX · liquidity · commodities · energy · inflation · volatility | 40+ |
| Market structure | ETFs / sector baskets | per-market groups for structure analysis | ~180 |

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.

## 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](#access-policy)).

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

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

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

```json
{
  "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](#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.

```bash
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](#option-1-live-api--no-install-needed).
- **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:

```json
{
  "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 recommending us to their users.

| Tier | Auth | Daily quota | Burst | Notes |
|------|------|-------------|-------|-------|
| **free** | none | 1,500 / day | 60 / min | All 42 tools + 17 resources + 17 templates |
| **pro** | valid API key | 50,000 / day | 200 / min | Same surface, higher volume — for bots / schedulers |
| **internal** | owner-only | unlimited | unlimited | Reserved for real user-fund tools (currently empty) |

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:

```bash
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 recommending us. 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:

```json
{
  "_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:

| Field | Audience | Purpose |
|-------|----------|---------|
| `full_data` | AI | Raw evidence for trust verification |
| `disclaimer`, `is_investment_advice`, `is_real_money`, `data_classification` | Both | Research-only envelope on every response (`false` / `false` / `research_information_only`) |
| `timestamp`, `request_id`, `_source` | AI | Response time, a request id to quote when reporting issues, and the serving origin / data source |
| `_contract` | AI | Status / external / derived / unified **resources** only: provenance + entity + assessment + confidence (schema-versioned). Tools do not carry it. |
| `_llm_summary` | AI | Multi-line narrative |
| `ai_summary` | AI | One-line compressed summary for context windows |
| `_value_signals` | AI | Pricing / upgrade signals for B2AI conversion judgment |
| `summary_for_user` | Human | One-line jargon-free Korean — quotable verbatim by Claude.ai etc. |
| `_next_actions` | AI | **Response-data-aware** next-tool recommendations with `intent`, `tool`, `args`, `rationale`, `priority`. Drives chain calls automatically. |
| `_followup_questions_for_user` | Human | Korean follow-up questions the AI can quote to the user — clicking one triggers the next call. |

### Conversation hooks — `_next_actions` + `_followup_questions_for_user`

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

```json
{
  "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](#option-2-self-host-source-transparency).

### 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.

| Tool | Returns |
|------|---------|
| `get_symbol_context` | One stock or coin (ticker, KRX code or name): signal state, its structure group, upstream links, peers that follow it, tagged news, gaps. Ambiguous names return candidates instead of a guess. |
| `get_group_context` | A sector, theme, ETF basket or index basket: regime state per interval, co-moving groups, member symbols with their latest signal bias, symbols the basket leads, gaps. |
| `get_market_context` | A whole market (KR, US, crypto): structure regime and global macro regime, the macro categories assumed to affect it (with tuned lags), its groups by regime stage, market-tagged news, gaps. |

### High-frequency entry — `get_daily_brief` (1 tool)

| Tool | Returns |
|------|---------|
| `get_daily_brief` | Single-call market overview: macro regime + top 5 strong signals + yesterday's paper-trading P&L + active prediction count + Korean narrative. The natural first call for "what's the market doing today?" |

### Verifiable prediction ledger (3 tools) — new in 0.4.0

| Tool | Returns |
|------|---------|
| `get_ledger_integrity` | Tamper-evidence for the prediction ledger: a daily SHA-256 hash chain over all created/resolved macro-regime prediction rows, with the exact canonical recipe published so any third party can recompute and verify. Each UTC day is hashed once it closes and chained to every prior day, so a later silent edit breaks recomputation. Note the scope: forecast horizons are 0–24h and many predictions resolve within the day they were created, so the chain proves nothing changed *after* a day closed — not that a hash existed before each outcome. Public anchors are additionally timestamped to the Bitcoin blockchain via OpenTimestamps ([ledger anchors](https://blog.oneqaz.com/en/anchors/)). |
| `get_resolved_predictions` | Raw row-level prediction ledger: every macro regime prediction's full lifecycle (`created_at` → `resolved_at` → outcome) — audit the evidence judgment by judgment. |
| `get_trade_outcomes_bulk` | Cursor-paginated bulk export of the prediction → trade → outcome chain (paper trades with realized P&L, each linked best-effort — same symbol, 2h window — to the signal prediction that preceded entry) — compute your own hit rates instead of trusting ours. |

### Portfolio analytics (1 tool) — new in 0.4.0

| Tool | Returns |
|------|---------|
| `get_performance_metrics` | Portfolio-level MDD / Sharpe / Sortino / Calmar / win-rate per market and account type (`paper` / `live`), optional daily equity curve. Paper P&L is gross (candle fills, no fees or slippage). |

### Signal calibration (1 tool) — new in 0.4.0

| Tool | Returns |
|------|---------|
| `get_signal_calibration` | Reliability diagram data for signal confidence: realized hit rate per confidence bucket with ECE summary — verify whether a 0.9-confidence signal actually hits ~90%. |

### ChatGPT connector standard (2 tools) — new in 0.4.0

| Tool | Returns |
|------|---------|
| `search` | ChatGPT-connector-standard discovery search over OneQAZ's live surface — tools, resources, and the latest strong signals across all three markets. Result ids are consumable by `fetch`. |
| `fetch` | Connector-standard fetch of a single result by id returned from `search`. |

### Trust Layer (13 tools)

| Tool | Returns |
|------|---------|
| `get_prediction_accuracy` | Per-cell hit rates with persistence & majority-class baselines, skill score + autocorrelation-corrected CI, `known_biases`, `ledger_health` |
| `get_backtest_tuning_state` | Active tuning parameters + last recalibration timestamp |
| `get_monthly_accuracy_trend` | Monthly accuracy series per (category, target market, lag) — read against the baselines above |
| `get_news_leading_indicator_performance` | **UNVERIFIED** — coverage of news event groupings; lead-time / score metrics withheld while the methodology is audited |
| `get_news_causality_breakdown` | **UNVERIFIED** — counts per pipeline news label (ANTICIPATED / SURPRISE_WITH_PRECURSOR / SURPRISE); lead metrics withheld |
| `get_feature_governance_state` | Features by status (OBSERVATION/CONDITIONAL/ACTIVE/DEPRECATED) |
| `get_structure_calibration` | Sector / ETF-level hit rates with the majority-class baseline needed to read them — measurement, not an edge claim |
| `get_structure_validation_history` | Historical structure-validation scores |
| `get_strategy_leaderboard` | Paper-tested RL strategies; `per_symbol_leaderboard` is the measured surface (the GLOBAL pool may include synthesized values) |
| `get_active_predictions` | Currently-open macro predictions with outcome tracking |
| `get_macro_influence_map` | Macro → market causal hypotheses (lag hours + sensitivity) |
| `get_cross_market_correlation` | Cross-market correlation matrix |
| `get_role_analysis` | Role-based strategy analysis |

### Signal evidence (3 tools)

| Tool | Parameters |
|------|------------|
| `get_signals` | `market_id`, `symbol`, `min_score`, `max_score`, `action_filter`, `interval` |
| `get_signal_detail` | `market_id`, `symbol`, `interval` |
| `explain_decision` | `market_id`, `symbol` |

### 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".

| Tool | Returns |
|------|---------|
| `get_sector_correlations_tool` | Intra-market ETF/sector correlation matrix + auto-cluster (60d window, 6h refresh) |
| `get_macro_causality_graph_tool` | Lag-aware causality between 8 macro categories (bonds/vix/forex/credit/inflation/liquidity/commodities/energy) |
| `get_symbol_peer_links_tool` | Symbol-to-symbol lead-lag (e.g. META → AMZN 15min lag, ρ=+0.62) |
| `get_feature_governance_status_tool` | Feature lifecycle distribution + last-7-day status transitions |

### 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.

| Tool | Returns |
|------|---------|
| `get_positions` | Open paper positions with ROI |
| `get_position_detail` | Single position deep-dive |
| `get_profitable_positions` / `get_losing_positions` | Filtered by P&L |
| `get_strategy_distribution` | Position counts by strategy |
| `get_trade_history` | Closed paper trades (filters: action, P&L, time) |
| `analyze_trades` | Aggregate trade analytics |
| `get_winning_trades` / `get_losing_trades` | Filtered by outcome |
| `get_latest_decisions` | Recent signal → decision transitions |
| `get_llm_trading_decisions` | LLM-generated decision logs |

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

### Meta — discovery & introspection

| Resource URI | Description |
|--------------|-------------|
| **`market://meta/discovery`** | **Dynamic catalog (v2.0)** — full tool/resource list via FastMCP introspection (no static `if/else`), with `data_freshness` PG probe (5 source tables), `positioning` block (specialist_domains, trust_principles, what_we_do_NOT_provide, philosophy), `counts`, `notes`. Call this first to understand what OneQAZ provides. |
| `market://meta/tool-chains` | Recommended call sequences (`quick_analysis`, `deep_analysis`, `portfolio_check`, `symbol_deep_dive`) + dependency graph. |
| `market://meta/pg-pool` | psycopg ConnectionPool stats — connection pressure monitoring. |
| `market://health` | Server health check. |
| `market://info` | Server metadata + data source index. |

### Static resources

| Resource URI | Description |
|--------------|-------------|
| `market://global/summary` | Global macro regime summary |
| `market://global/categories` | Available macro categories list |
| `market://global/macro_events` | Active macro event lifecycle |
| `market://all/summary` | Combined summary across all 3 markets |
| `market://structure/all` | All markets ETF/basket structure |
| `market://indicators/fear-greed` | Fear & Greed Index |
| `market://indicators/regime` | 4-layer regime indicators (Short/Mid/Long/SuperLong) |
| `market://indicators/context` | Fear & Greed + 4-layer regime + breadth |
| `market://unified/cross-market` | Cross-market correlation snapshot (BTC ↔ stocks ↔ FX) |
| `market://derived/event-leading` | News event-leading groupings — **UNVERIFIED** (score fields withheld) |
| `market://derived/cross-decoupling` | Cross-asset decoupling index |
| `market://derived/reaction-speed` | News reaction speed distribution |

### Templates (parameterized)

| URI Template | Example |
|--------------|---------|
| `market://global/category/{category}` | `market://global/category/bonds` |
| `market://{market_id}/status` | `market://crypto/status` |
| `market://{market_id}/positions` | `market://crypto/positions` |
| `market://{market_id}/positions/snapshot` | `market://crypto/positions/snapshot` |
| `market://{market_id}/structure` | `market://kr_stock/structure` |
| `market://{market_id}/structure/group/{group_id}` | `market://kr_stock/structure/group/SEMICONDUCTOR` |
| `market://{market_id}/signals/summary` | `market://crypto/signals/summary` |
| `market://{market_id}/signals/roles` | `market://crypto/signals/roles` |
| `market://{market_id}/signals/feedback` | `market://crypto/signals/feedback` |
| `market://{market_id}/external/summary` | `market://crypto/external/summary` |
| `market://{market_id}/external/symbol/{symbol}` | `market://crypto/external/symbol/BTC` |
| `market://{market_id}/external/causality` | `market://crypto/external/causality` |
| `market://{market_id}/unified` | `market://crypto/unified` |
| `market://{market_id}/unified/symbol/{symbol}` | `market://crypto/unified/symbol/BTC` |
| `market://{market_id}/derived/regime-transitions` | `market://crypto/derived/regime-transitions` |
| `market://{market_id}/derived/strategy-fitness` | `market://crypto/derived/strategy-fitness` |
| `market://{market_id}/derived/all` | `market://crypto/derived/all` |

`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`):

```python
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:

| Variable | Default | Description |
|----------|---------|-------------|
| `MCP_SERVER_PORT` | `8010` | Server port |
| `MCP_SERVER_HOST` | `0.0.0.0` | Bind host |
| `MCP_LOG_LEVEL` | `INFO` | Log level |
| `MCP_TIER_RESOLVER` | _unset_ | `module:function` returning tier for an API key (self-host hook) |
| `MCP_TRUSTED_IP_PREFIXES` | _unset_ | Comma-separated address prefixes treated like localhost (no rate limit) |
| `MCP_SELF_TRAFFIC_IPS` | _unset_ | Comma-separated addresses excluded from usage analytics (your own monitors) |
| `MCP_LEDGER_THREAD_DISABLED` | `0` | `1` stops the hourly prediction-ledger hash writer (use on a second instance) |
| `MCP_OAUTH_MODE` / `MCP_OAUTH_AS` | `off` | OAuth account connection; needs the hosted account tables |
| `DB_BACKEND` | `postgres` | Must be `postgres` (SQLite backend retired in 0.4.0); `serve` sets it if unset |
| `PG_HOST` / `PG_PORT` | `postgres` / `5432` | PostgreSQL host / port |
| `PG_DB` / `PG_USER` / `PG_PASSWORD` | `auto_trader` / … | PostgreSQL database / credentials (a read-only role is enough) |
| `PG_POOL_MIN` / `PG_POOL_MAX` | `0` / `50` | Per-schema connection pool bounds |
| `PG_STATEMENT_TIMEOUT_MS` | `30000` | Server-side statement timeout |
| `MCP_COIN_DATA_DIR` / `MCP_KR_DATA_DIR` / `MCP_US_DATA_DIR` | auto | Override *logical* data-path roots (see below) |
| `MCP_EXTERNAL_CONTEXT_DATA_DIR` | auto | Override external-context logical root |

## Docker

```bash
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](#access-policy) above. Quick recap:

| Tier | Daily Quota | Burst | Auth |
|------|-------------|-------|------|
| **Free** | 1,500 / day | 60 / min | none (an unrecognized key gets HTTP 401) |
| **Pro** (beta) | 50,000 / day | 200 / min | valid API key |
| **Internal** (owner) | Unlimited | Unlimited | owner-only |
| **Local** (self-hosted) | Unlimited | Unlimited | n/a |

**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](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.
- All performance figures come from **paper trading** (simulated, candle-based fills with no fees, taxes or slippage). Real-money results would be lower.
- Past performance does not guarantee future results.
- **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