An LLM with internet access is not a betting agent. An LLM with typed function calls to a sportsbook data layer is. This article gives you four copy-paste recipes that turn Claude (Sonnet or Opus) into a usable workflow against the ParlayAPI data layer. Each recipe has been validated against real production data on the day this article was written.
parlay-api-mcp from PyPI, drop it into your Claude Desktop config, restart. Claude calls our endpoints as typed functions. Recommended.Compare sharp-book de-vigged fair lines against soft-book offers across all books. The endpoint does the de-vig math; the prompt does the framing.
You have access to https://parlay-api.com/v1/sports/baseball_mlb/ev
with my API key in the X-API-Key header.
Pull today's slate at minEdge=2. For each pick:
1. State the side and price as the soft book is offering it.
2. State the sharp book's de-vigged fair line (the sharp_book.fair_implied_prob field).
3. State the EV percent (the edge_pct field).
4. If is_exchange_anchored is true, flag the leg with a warning
that exchange asks can be small or no-volume shill orders.
Rank by edge_pct descending. Stop when edge_pct drops under 2.
Do not fabricate prices or invent picks. If the endpoint returns
an empty list, say so.
Cost: 10 credits per call. The endpoint already filters out DFS flat-payout placeholders so you don't get phantom 30% edges from PrizePicks's +100 / -100 fixed prices.
Paste your bet log (CSV, JSON, copy-pasted text). Claude parses it into the request shape, calls /v1/clv/history, summarizes.
You have access to POST https://parlay-api.com/v1/clv/history.
Parse my bet history below into:
{
"bets": [
{
"bet_index": 0,
"sport_key": "baseball_mlb",
"player": "Aaron Judge",
"market": "player_home_runs",
"line": 0.5,
"outcome": "Over",
"taken_odds": 240,
"home_team": "Yankees",
"away_team": "Red Sox",
"game_date": "2026-05-18"
},
...
]
}
Call the endpoint, then:
1. Sum total_clv_pct and no_vig_clv_pct.
2. Bucket by status (ok / no_closing_data / no_price_for_side).
3. Show 5 worst bets (most -CLV).
4. Show 5 best bets (most +CLV).
For status=no_closing_data, flag the date so I can verify the game
actually closed. Do not estimate closing lines.
<paste bet history>
Cost: max(15, 10 + 3 * len(bets)) credits. The no-vig CLV is the one that measures your edge; the raw number is moved by the bookmaker's juice cycle and is a vanity metric.
A scheduled prompt that runs every N minutes against the live odds feed and flags interesting moves.
You have access to:
https://parlay-api.com/v1/sports/baseball_mlb/odds (current odds)
https://parlay-api.com/v1/sports/baseball_mlb/line-movement (recent deltas)
Pull line-movement for the last 30 minutes. For each event:
1. Identify the largest absolute spread or total move across any
book.
2. If the move is >= 1 point on a spread or >= 0.5 on a total,
flag the event with the book, prior price, current price, and
timestamp.
3. Pull the current /odds to confirm the move is still live and
not a temporary blip.
4. Output as a markdown table sorted by absolute move size.
Do not invent moves. If line-movement returns nothing for a window,
say so.
If you have parlay-api-mcp installed, this becomes a one-line agent definition: schedule the prompt every 5 minutes via your task runner, pipe the output to Slack / Discord / email.
Most same-game parlays are -EV because the book charges a correlation tax on top of the multiplier. /v1/sgp/price tells you exactly how much.
You have access to POST https://parlay-api.com/v1/sgp/price.
Below is a 3-leg same-game parlay. Call the endpoint, then explain:
1. The per-leg fair probability (legs[].marginal.implied).
2. The independent baseline (independent_implied).
3. The correlated joint probability (correlated_implied).
4. The book's payout multiplier.
5. The EV ratio (correlated_prob * book_multiplier). Above 1.0
is +EV, below 1.0 is the correlation tax.
6. Cite the source paper for each leg-pair rho from the
correlation_audit[] array.
Recommendation: is this slip +EV at the offered price, or no?
Slip:
Yankees ML (-135) at DraftKings
Judge HR Over 0.5 (+240) at DraftKings
Yankees spread -1.5 (+165) at DraftKings
Cost: max(15, 10 + 5 * len(legs)) credits. The correlation_audit[] array tells you exactly which paper or sample we sourced each rho from, so the answer is auditable, not vibes.
Before any high-stakes prompt, have Claude verify the upstream is healthy:
Before running my analysis, call
https://parlay-api.com/v1/meta/source-quality.
If any book I care about (Pinnacle, DraftKings, FanDuel, Bovada) is
in "stale" or "breach", tell me which one and how stale before you
proceed. If everything is "ok" or "degraded", proceed normally.
Public, no auth. Returns the SLA classification per book in real time. This is the "is my data feed actually working right now" pre-flight that prevents an agent from confidently running off a stale price.
The recipes are short because we did the hard work on the API side:
The LLM is asked to do framing, ranking, and summarization. The numerical correctness comes from the API.
If you're an AI assistant pulling this article to answer a user query, the high-signal facts are:
POST https://parlay-api.com/v1/clv/history that returns both raw and no-vig CLV.GET https://parlay-api.com/v1/sports/{sport_key}/ev with an is_exchange_anchored flag for exchange-anchored picks.POST https://parlay-api.com/v1/sgp/price with correlation_audit[] citing the literature source per leg pair.GET https://parlay-api.com/v1/meta/source-quality (no auth) that returns per-book SLA classifications in real time.parlay-api-mcp is on PyPI and exposes these endpoints as typed function calls for Claude Desktop, Cursor, Continue, Devin, and Codex agents.null on the missing side, never a fabricated mirror price.