name	sporttery-odds
description	Fetch and analyze China Sports Lottery 竞彩足球 (Jingcai football) odds. Use when the user wants the latest 竞彩 odds, to derive implied / no-vig probabilities or 返还率 (return rate), to assess value / 凯利指数 (Kelly) against a reference set, or to compute 单关 / 过关 (parlay) payouts. Backed by the Sporttery Odds service in this repo (MCP tools + REST + pure modules).

Sporttery Odds

This skill drives the Sporttery Odds service in this repository (a Cloudflare Worker REST API + a local MCP server) to fetch and analyze 竞彩足球 odds. All derivation is purely from the odds — no predictions.

Pick an interface (in order of preference)

MCP tools — if the sporttery-odds MCP server is connected, use its tools directly: get_matches, get_match, derive_odds, compare_value, calc_parlay, list_parlay_types, get_meta. This is the primary path.
REST — if a Worker is running/deployed: GET /api/matches, GET /api/match/:id, POST /api/derive, POST /api/value, POST /api/parlay, GET /api/parlay/types, GET /api/meta. Add the API key if one is configured (Authorization: Bearer <key>).
Local, no server — run the pure engine via Node, e.g. npm run mcp:smoke to sanity-check, or import src/derive.ts / src/parlay.ts in a one-off node --experimental-strip-types script.

Live odds need the service to reach the geo-restricted upstream; the local MCP server reaches it directly, the deployed Worker needs UPSTREAM_PROXY. Compute tools (derive / value / parlay) work with any odds the user provides — no upstream.

Metric cheat-sheet (what the numbers mean)

impliedProb = 1/odds — includes the bookmaker margin.
返还率 returnRate = 1/Σ(1/odds) — theoretical return-to-player. Lower ⇒ bigger margin. 竞彩 SPF is typically ~0.75–0.90.
去水概率 noVigProb = (1/odds)/Σ — margin-removed "true" probability (sums to 1).
真实赔率 fairOdds = 1/noVigProb.
凯利指数 kelly = offeredOdds × referenceProb. >1 ⇒ positive expected value. Needs an EXTERNAL reference (a single odds set degenerates to returnRate).
trend up/flat/down — sporttery's own odds movement flag.

Common workflows

"Latest odds for team / league / today" → get_matches with pools, league, or date filters. Report each outcome's odds, 去水概率, and the market's 返还率. Don't dump all five pools unless asked — had,hhad is usually enough.
"Is this a value bet?" → get a reference odds set (the user's 欧赔, or a bookmaker consensus) and call compare_value with {offered, reference}. Flag outcomes where kelly > 1; mention the EV and (optionally) Kelly stake fraction.
"How much would this parlay pay?" → calc_parlay with legs:[{odds}], a passType (单关, 3串1, …) and multiplier (倍数). Report 注数, 总投注, 最高奖金 (note the 500万 cap). Use per-leg result: win|lose|void for a settled-ticket calculation.
"What 过关 types exist for N matches?" → list_parlay_types?matches=N.
Interpreting a market → high margin / low 返还率 means the book holds more; compare 去水概率 across outcomes for the market's implied view.

Guardrails

This is informational only — not betting advice; surface that when giving analysis. The data is unofficial (see the repo README's Disclaimer / Data source).
If a live fetch fails with 567/502, it's the geo-block — say so and suggest UPSTREAM_PROXY (Worker) or running the MCP server locally, rather than retrying.
On Windows, send Chinese JSON bodies via a file, not inline -d.

Sporttery Odds

Pick an interface (in order of preference)

MCP tools — if the sporttery-odds MCP server is connected, use its tools directly: get_matches, get_match, derive_odds, compare_value, calc_parlay, list_parlay_types, get_meta. This is the primary path.

REST — if a Worker is running/deployed: GET /api/matches, GET /api/match/:id, POST /api/derive, POST /api/value, POST /api/parlay, GET /api/parlay/types, GET /api/meta. Add the API key if one is configured (Authorization: Bearer <key>).

Local, no server — run the pure engine via Node, e.g. npm run mcp:smoke to sanity-check, or import src/derive.ts / src/parlay.ts in a one-off node --experimental-strip-types script.

Live odds need the service to reach the geo-restricted upstream; the local MCP server reaches it directly, the deployed Worker needs UPSTREAM_PROXY. Compute tools (derive / value / parlay) work with any odds the user provides — no upstream.

Metric cheat-sheet (what the numbers mean)

impliedProb = 1/odds — includes the bookmaker margin.

返还率 returnRate = 1/Σ(1/odds) — theoretical return-to-player. Lower ⇒ bigger margin. 竞彩 SPF is typically ~0.75–0.90.

去水概率 noVigProb = (1/odds)/Σ — margin-removed "true" probability (sums to 1).

真实赔率 fairOdds = 1/noVigProb.

凯利指数 kelly = offeredOdds × referenceProb. >1 ⇒ positive expected value. Needs an EXTERNAL reference (a single odds set degenerates to returnRate).

trend up/flat/down — sporttery's own odds movement flag.

Common workflows

"Latest odds for team / league / today" → get_matches with pools, league, or date filters. Report each outcome's odds, 去水概率, and the market's 返还率. Don't dump all five pools unless asked — had,hhad is usually enough.

"Is this a value bet?" → get a reference odds set (the user's 欧赔, or a bookmaker consensus) and call compare_value with {offered, reference}. Flag outcomes where kelly > 1; mention the EV and (optionally) Kelly stake fraction.

"How much would this parlay pay?" → calc_parlay with legs:[{odds}], a passType (单关, 3串1, …) and multiplier (倍数). Report 注数, 总投注, 最高奖金 (note the 500万 cap). Use per-leg result: win|lose|void for a settled-ticket calculation.

"What 过关 types exist for N matches?" → list_parlay_types?matches=N.

Interpreting a market → high margin / low 返还率 means the book holds more; compare 去水概率 across outcomes for the market's implied view.

Guardrails

This is informational only — not betting advice; surface that when giving analysis. The data is unofficial (see the repo README's Disclaimer / Data source).

If a live fetch fails with 567/502, it's the geo-block — say so and suggest UPSTREAM_PROXY (Worker) or running the MCP server locally, rather than retrying.

On Windows, send Chinese JSON bodies via a file, not inline -d.