TKB Picks MCP
MCP server that wraps SportsGameOdds and BALLDONTLIE APIs to provide sports scheduling, odds, injuries, and team statistics tools for building TKB Picks threads.
README
TKB Picks MCP Server
Wraps SportsGameOdds (odds/schedule/stats) and BALLDONTLIE (injuries) into MCP tools for TKB Picks thread-building.
This round: response-size optimization (root-cause fix for the earlier OOM crash)
What changed: Every tool that fetches odds now requests the EXACT oddID(s) it needs directly from SGO's API, instead of pulling the full event (which can contain 1000+ markets) and filtering client-side afterward. Confirmed via SGO's own FAQ and docs that request-level filtering (oddIDs, playerID, bookmakerID params) is the correct, intended way to keep responses small - this was the real root cause of the earlier out-of-memory crash, more than raw traffic volume.
Files changed:
sgoClient.ts- addedbookmakerIDandincludeOpposingOddsto the request params (bookmakerID is wired through; includeOpposingOdds is available for future use but not yet used by any tool)odds.ts- rewritten sooddIDsare built BEFORE the event lookup and passed on the very first request, even in the teamName-lookup case (since statID/entity/side are deterministic ahead of time, we never need to guess-then-filter)yesNoProps.ts,periodOdds.ts- same fix: the already-knownoddIDis now passed via theoddIDsparam on the fetch instead of being filtered out of the full response after the facthitRateAggregator.ts,splitsAggregator.ts- these never needed odds data at all (onlyresults/teams.score), so they now pass a minimal single dummyoddIDsvalue to shrink the odds payload to near-zero instead of pulling the full odds object for every game in the lookback window
New param: tkb_get_odds now accepts optional preferredBookmakers (comma-separated, e.g. "fanduel,draftkings") to control which book's price is returned, instead of whatever happened to come back first.
Practical effect: faster responses, and this should meaningfully reduce (likely eliminate) the memory pressure that caused the earlier "Ran out of memory (used over 512MB)" crashes - independent of the Render instance upgrade, which was a workaround for the same underlying issue. Nothing about the tool names, parameters (aside from the one addition), or how threads get built has changed - this is purely a request-efficiency fix.
Tools (10)
tkb_get_schedule- game schedule, date/team/conference filteringtkb_get_odds- moneyline/spread/total, or an individual player's over/under prop (now requests exact oddIDs directly)tkb_get_player_hit_rate- real recent-game-log hit-rate check (now requests minimal odds payload)tkb_get_injuries- structured injury status from BALLDONTLIE (team field confirmed fixed:player.team.display_name)tkb_get_team_split- home/road/opponent-specific record (now requests minimal odds payload)tkb_get_team_record- overall record from SGO's standings datatkb_get_yes_no_prop- milestone-style bets (now requests exact oddID directly)tkb_get_period_odds- period-specific lines (now requests exact oddID directly)tkb_debug_raw_event- dumps raw SGO event JSONtkb_debug_raw_injuries- dumps raw BALLDONTLIE injuries JSON
Confirmed working / fixed across all prior rounds (still in effect)
- Start time field (
status.startsAt) - Schedule/odds teamName searches properly date-bounded
- Moneyline/spread odds correct with or without explicit
side - Hit-rate and team-split date bounding (no more stale/ancient data)
- BALLDONTLIE injuries team field (
player.team.display_name) - No
lineupsfield exists on SGO events - starting pitcher confirmation is a permanent live-search requirement, not something this API can provide
Still genuinely unverified
- Player
teamIDupdate speed after a real trade (though one real offseason trade - Semien to Mets - was confirmed correctly reflected) includeOpposingOddsbehavior (added to client, not yet used by any tool - worth adopting in a future round to cut requests further for two-sided markets)
Operational notes
Render instance: currently on Standard (2GB RAM). This round's fix targets the root cause of the earlier OOM crashes directly, so the bigger instance should now have even more headroom than before.
MCP connector caching: remove/re-add if new tools/behavior don't show up after redeploy.
Start Command must be exactly npm start.
Deploy to Render
- Upload this folder's contents to the GitHub repo (overwrite existing files)
- Render auto-redeploys
- Verify:
https://YOUR-SERVICE.onrender.com/health - Remove/re-add the MCP connector if tools seem stale
Adding NBA/NHL later
- Get a BALLDONTLIE subscription for that sport
- Add one entry to
SPORT_CONFIGinsrc/constants.ts - Push to GitHub
Local development
npm install
npm run build
SGO_API_KEY=xxx BDL_API_KEY=xxx npm start
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。