Native AI integration

Chat with Predixsport data
directly in Claude

Ask questions about NBA, football and tennis predictions in plain language. Claude pulls live data from our neural networks and answers in context.
Setup takes about four minutes. No code.

Sign in to begin ↓

What this is

Predixsport ships a Claude connector — an integration that lets Anthropic's Claude (the conversational AI) call our prediction models on demand. Once it's connected to your Claude account, you can ask anything you would normally ask of a sports analyst: which matches look interesting this weekend, how confident a forecast is, how a specific model has performed historically. Claude returns answers grounded in our data, not in generic web knowledge.

Every number Claude shows comes from the same neural networks that power predixsport.com. Predictions update daily; calibration is documented and queryable.

What you can ask

Four representative questions, written exactly as you would ask them in chat:

"What Serie A matches this weekend look most interesting? Pick one and explain why."

🎾

"Compare Ruud and Khachanov's upcoming match — win probability, total games, tiebreak chance."

🏀

"How well-calibrated is your NBA winner model? Can I trust a 70% prediction?"

📊

"Show me Napoli's prediction history over the last month. Anything that stands out?"

Claude picks the right tool, calls our API, and answers in plain language. You don't pick endpoints, write queries, or parse JSON.

A Claude conversation: the user asks about upcoming NBA matches, and Claude returns a structured prediction card for Detroit Pistons vs Cleveland Cavaliers with 70% / 30% win probability, spread DET -6.2, total 214.8, and projected starters with points/assists/rebounds
A real conversation. The user asks; Claude calls the Predixsport tools and renders a structured prediction card — win probabilities, spread, total, projected starters — in one turn.

Who this is for

  • Sports analytics enthusiasts and quantitative fans — people who care about why a forecast says what it says, and want to interrogate the model rather than just read its output.
  • Sports journalists, broadcasters and content creators — faster sourcing of probability-backed angles for previews, recaps and podcasts, with the model's working clearly attributable.
  • Fantasy players and data-curious supporters — anyone who wants a deeper read on form, calibration and player projections than the conventional table view gives.

Why this and not something else

Reasonable question. We offer three ways to use the same data — the connector below, a REST API for code, and a JSON/CSV download on every match page. Here's how to think about choosing.

Why not just ask Claude to read your website?

You can, and it will sort-of work. The website is rendered for humans: probabilities are rounded to one decimal, distribution charts are images, and the calibration data lives on a different page. Stitching all of that together by parsing HTML and pixels is exactly what browser-based agents do badly. The connector gives Claude the raw probability arrays, the model identifier and the calibration history in one tool call — so it can answer follow-ups the scraping route can't.

Why not just download the JSON or CSV?

The download is a snapshot at the moment you click. NBA inference in particular only runs a few hours before tip-off, so a file you grabbed in the morning may show no matches when predictions actually publish. The connector pulls fresh data on every question. It also opens the door to scheduled prompts (Claude Pro and Max plans) — a recurring question that runs on its own without you clicking anything.

How is this different from your REST API?

The connector runs on top of the API. Same endpoints, same data. The API is for code — Python, dashboards, model stacks, automated pipelines. The connector is for people who live inside Claude and don't want to write that integration: no SDK, no rate-limit handler, no JSON parsing. The cost is that you are tied to Claude's interface, which is the point if Claude is already where your other work happens.

What does MCP let me do that nothing else does?

One conversation, multiple data sources. With other MCP servers connected to your Claude account — a fantasy roster source, a personal notes app — you can cross-reference all of them in a single back-and-forth without opening three apps. That is what the protocol was designed for, and it is the part that's awkward with any of the other layers.

Longer take with concrete examples on the blog: Predixsport with Claude: connector, API, or download — which fits you.

Where it works

You set up Predixsport once and it works everywhere you sign in to Claude. Connectors are stored on your Claude account, not on your device — once added (most easily from the web app at claude.ai or the desktop app), they appear automatically on the iOS and Android apps too. There is no separate mobile installation.

You only set up once. Pick web OR desktop — whichever is easier — do it once, and Predixsport propagates to every other Claude client signed in to that account.

Platform Can you set up here? Will it work once added elsewhere?
Claude on the web (claude.ai) Yes — easiest, recommended Yes
Claude desktop — macOS & Windows (same procedure) Yes — equivalent to the web Yes
Claude mobile — iOS & Android — (no Add-Connector UI on mobile) Yes — inherited from your account
ChatGPT (desktop & mobile) Roadmap — coming in the months ahead

Plans: all Claude subscription tiers can use custom connectors. Anthropic limits the free tier to one custom connector per account; Pro, Max, Team and Enterprise have no documented cap. Plan tiers also differ in Claude's overall usage limits — see claude.com/pricing. Predixsport does not add plan restrictions of its own.

How the flow works

You type a question in Claude. The connector translates it into a call to our prediction service, which returns the data, and Claude composes the natural-language answer. You see only the question and the answer; the steps in between are automatic.

1 Claude You ask a question in chat.
2 Predixsport connector Selects the tool that fits your question.
3 Predixsport models Return probabilistic predictions and metadata.
4 Claude Replies in plain language with sources.

One-time setup — three minutes, three steps

You only do this once. The connector is saved on your Claude account — not on your device. After the first setup below, Predixsport is automatically available on every Claude client where you're signed in: web, desktop (macOS & Windows), iOS, Android. Nothing to repeat per device, nothing to reinstall.

Predixsport uses standard sign-in — there are no API keys to copy. Add the connector in Claude, click Authorize on the Predixsport screen that opens, and you're ready to ask questions.

1

Add Predixsport as a connector in Claude

Do this once on the web or desktop app — it will then work on all your Claude clients, including mobile, automatically (the connector lives on your Claude account, not on your device). On claude.ai or the Claude desktop app:

  1. Open Settings → Connectors in the sidebar.
  2. Click + then Add custom connector.
  3. Fill in two fields — leave Advanced settings empty:
  • Name Predixsport — this is the label Claude will display in the connector list and inside chats. Please keep it as Predixsport so it matches the rest of this guide and our support docs.
  • Server URL https://mcp.predixsport.com/sse

Click Add. Claude opens a Predixsport sign-in window. The procedure is identical on macOS and Windows desktop apps — same menu name, same fields.

Note on Advanced settings. The OAuth Client ID and OAuth Client Secret fields are optional — Predixsport registers your Claude installation automatically. Leave both empty.

2

Sign in and authorize on Predixsport

Right after you click Add, Claude opens a window on predixsport.com:

  1. If you don't already have a Predixsport account, sign in with Google — same Google account you use to read articles on predixsport.com. New accounts start on the Free tier (200 calls/month).
  2. You'll then see a Predixsport Authorize Claude screen listing what Claude can access (predictions, ratings, performance metrics) and your current monthly quota.
  3. Click Authorize. The window closes and Claude lists the Predixsport tools it can now call.

On mobile? The iOS and Android apps don't expose an "Add connector" UI — but they don't need to. The Predixsport connector you just authorized appears in mobile Claude automatically the next time you open the app, because it travels with your Claude account, not your device.

Claude tool permissions panel listing the nine Predixsport tools (get_upcoming_predictions, search_predictions, get_match_detail, get_nba_player_props, get_current_ratings, get_model_performance, get_model_calibration, list_sports, list_models)
After authorization succeeds, Claude lists the nine Predixsport tools it can call — predictions, ratings, performance, calibration and catalog.
3

Start asking questions

Open a new chat and ask anything in plain language. Claude surfaces the Predixsport tools in its Search and tools menu; you don't need to select them explicitly — the model picks the right one for each question.

If Claude doesn't seem to see the connector after you add it, fully quit and reopen the app or refresh the browser tab.

A follow-up question in the same Claude conversation: the user asks for probability distributions of points by player, and Claude renders an interactive Poisson histogram for Cade Cunningham (expected 24.9 pts, mode 23, std dev ±8.7, P(20+ pts) 71.1%) with tabs to switch player
Follow-ups unlock depth. Ask about probability distributions and Claude renders the full Poisson curve per player — mean, mode, std dev, threshold probabilities — not just an expected value.

What's covered today

  • NBA — regular season and playoffs. Match-level predictions (win probability, point spread, total points) and per-player projections (points, rebounds, assists, three-pointers made).
  • Football — Serie A, Premier League, La Liga, Bundesliga, Ligue 1. Match result (1X2), over/under 2.5 goals, both teams to score, expected goal spread, corners and shots.
  • Tennis — ATP. Winner probability, total games, game spread, exact set score, aces and double faults projections, tiebreak likelihood.

Every prediction is generated by a dedicated neural network and evaluated independently. Per-model calibration data is queryable through the connector — ask Claude "how well-calibrated is your <model_id>?" to see the reliability curve.

What's coming next

  • ChatGPT integration — in the months ahead, via OpenAI's connector framework. The underlying API is the same; the entry point changes.
  • WTA tennis — women's tour coverage with the same model family already deployed for ATP.
  • Rating histories — query Elo and Glicko ratings over time for teams and players.

Privacy

The connector only sees the tool calls Claude makes on your behalf — for example "fetch upcoming NBA matches" with the relevant parameters. It does not see your chat with Claude. Predixsport counts monthly usage tied to your account for rate limiting and abuse prevention. We do not store the content of your prompts.

Revoke your key at any time from the API access page. Full details in our privacy policy.

Troubleshooting

Claude doesn't see the Predixsport tools

Fully quit and restart Claude after adding the connector (Cmd+Q on macOS, the file menu's Quit on Windows; hard-reload the browser tab on the web; force-close and reopen the mobile app). Closing the window alone does not reload the connector configuration.

If after restart the tools still don't appear, open Settings → Connectors: any configuration error is displayed there alongside the entry.

Claude returns "Invalid or expired API key"

Verify the key matches what you copied from the API access page. A trailing space or accidental quote during paste is the most common cause. If unsure, generate a new key — the old one is revoked automatically when you regenerate from the same page.

The connector returns 503 Service Unavailable

A transient signal that our backend is briefly busy — typically during a daily model retrain. The Retry-After header tells Claude when to try again; it usually resolves in seconds. If it persists for more than five minutes, send us a message via the contact form.

I don't see any NBA matches in the results

NBA predictions are gated on the expected starting five and injury reports, which are published only a few hours before tip-off. An empty result for a wide date range is the normal state during the day; predictions appear on game day. Ask Claude for the next scheduled card and it will adjust the window for you.

Tennis predictions show a date that looks old

The date field on tennis responses is the tournament's scheduled start day, not the day a specific match is played. A Masters event that begins on day X has every match across all rounds tagged with date X, even matches played a week later. The data is fresh; only the field's meaning differs from football and basketball conventions.

Are there usage limits?

The free tier is 200 connector requests per calendar month, shared across all your active keys. That covers a few dozen Claude conversations with heavy back-and-forth, so most people never notice it. Hit the limit and the connector replies with a clear message telling you exactly when access returns (the 1st of the next month, UTC). Drop us a line at our contact form and we'll happily lift the cap. Anthropic plans separately set overall Claude usage; their limits are at claude.com/pricing.

For developers

If you'd rather drive the predictions from your own code, the REST API powers this connector — same endpoints, same data, same rate limits. Use it for Python notebooks, dashboards, automated pipelines, or anything where you want raw JSON over HTTPS instead of a chat interface. Full reference and key management on the API access page.

Ready to start?

Sign in, generate your access key, paste three fields into Claude and you're done. Four minutes from now you can ask Claude what looks interesting this weekend.

Sign in to begin →

Questions, feedback, or stuck on setup? Send us a message — we read every reply.