Batru is built agent-native: the same calibrated predictions the site runs on, exposed to LLM agents as a hosted MCP server and to code as a Python SDK. Read-only, no API key required.
AI assistants can pull live Dota 2, Deadlock and Marvel Rivals stats directly from Batru: the hosted MCP server at https://batru.gg/mcp exposes read-only tools for tier lists, counters, synergies, ban rates and calibrated win-probability predictions — no API key, no account. A machine-readable index lives at /llms.txt.
The hosted Batru MCP server gives any MCP-capable assistant (Claude, ChatGPT, and a growing list of clients) tools backed by our live production models. Current public coverage includes Dota 2, Deadlock and Marvel Rivals: win-rate prediction, pick recommendations, counters & synergy, tier lists (including per-rank), single matchups, hero momentum, ban rates and our public model-evaluation ledger. Every data tool takes a game parameter — one compact surface, three games. The server also exposes the standard search / fetch pair (so ChatGPT deep research can use it) and clickable prompts (draft_help, counter_check, meta_brief). The tools instruct the model to report our numbers verbatim instead of guessing from memory.
Hosted endpoint — no install, always the newest tool set. Paste it into Claude (Settings → Connectors → Add custom connector) or ChatGPT (walkthrough below), or:
https://batru.gg/mcp
Try it in 10 seconds — this button sends a real request to the same live model the MCP tools call (no key, straight from your browser):
POST https://batru.gg/api/winrate
{"radiant":["antimage","axe"],"dire":["lina","tinker"]}ChatGPT — step by step (Plus/Pro):
Batru, Server URL https://batru.gg/mcp, Authentication No Auth → acknowledge → Create → Connect.Read-only tools, no account or key on our side. Hosted calls are logged by default for usage analytics, quality, error diagnosis and abuse prevention; see our privacy policy. The same URL works in any MCP-capable client.
Claude Code — one line (hosted):
claude mcp add --transport http batru https://batru.gg/mcp
Prefer local stdio? Claude Desktop — claude_desktop_config.json(local stdio sends no MCP tool-call analytics; normal API request logs still apply, and the PyPI package may trail the hosted tool set):
{
"mcpServers": {
"batru": { "command": "uvx", "args": ["batru-mcp"] }
}
}Building something? Two recipes people actually use this for:
predict_winrate returns calibrated probabilities (a reported 60% is a real ~60% empirical win rate; the full evaluation ledger is get_model_track_record). Feed both models the same drafts and measure agreement — we make a solid reference line for evals. predict_winrate_batch scores a full-roster sweep (150 candidates) in one call; recommend_pick, get_counters and get_synergy cover the pick phase. Read-only; a free API key (below) raises rate limits. Endpoint docs & tool list · Listed on the official MCP Registry as gg.batru/batru-mcp · PyPI · GitHub
batru is the official Python client — a thin, read-only wrapper over the public prediction API. Its current client coverage is Dota 2, Deadlock and Marvel Rivals.
pip install batru import batru batru.winrate(radiant=["antimage", "axe"], dire=["lina", "tinker"]) # -> calibrated win probabilities from the live production model # (run "Try it in 10 seconds" above for the current numbers) batru.recommend(radiant=["antimage"], dire=["lina", "tinker"])
Agents don’t read vibes — they act on the number you hand them. Our probabilities are calibrated: when the model says 60%, that side wins about 60% of the time on held-out data. That property — not a headline accuracy figure — is what makes a prediction safe to build on. How we measure it.
Everything works without a key. If you’re building something that calls us often, grab a key — it lifts every limit 10× and gives your integration its own quota instead of sharing a per-IP one.
| Access | Prediction requests | How |
|---|---|---|
| No key | 60 / min per IP | Just call the API |
| With a key | 600 / min per key (each batch call = 1 request, up to 150 drafts) | Button below — instant, no approval |
| Bigger than that | Let’s talk | Contact us |
Requires an account — login exists only to stop abuse, same as the site.
Prefer files over APIs? Every Captains-Mode draft from TI 2025/2026 and EWC 2025/2026 — 538 games, 12,912 draft actions, plus archived pre-series model probabilities for TI 2026 — ships as Parquet/CSV under CC BY 4.0. Dataset page · Hugging Face.
Everything here is free and keyless, no strings attached. If you build something user-facing on this data, a small credit helps other players find the source — and it’s what keeps the API free:
Powered by <a href="https://batru.gg?utm_source=powered_by">batru.gg</a>
Agents get this for free: every MCP tool result already includes a ready-made citation / source_url field to show alongside the numbers.
The API behind these clients is free and keyless. Please keep request volume reasonable, let your client identify itself via User-Agent (the MCP server and SDK already do), and cache what you can. Building something that needs more — or a language we don’t cover yet? Talk to us.
Analysis only — these tools don’t provide betting advice, and we don’t endorse or facilitate sports betting.
Tier shifts, counters and hero pairings across the games Batru covers.
Unsubscribe anytime.