Batru — Competitive Game Intelligence for AI Agents

gg.batruPublished by batru.ggOfficial publisherv0.7.0Updated Oct 25, 2026

Live competitive-game predictions, meta data and decision tools for AI agents.

VerifiedStreamable HTTPWeb executableWeb Search & ScrapingData & Analytics

Overview

AI-generated overview

A read-only MCP server over batru.gg that gives an assistant live Dota 2, Deadlock and Marvel Rivals draft predictions, hero meta data and pro tournament stats.

What you can do
Look up and normalise hero names, or list the roster, for Dota 2, Deadlock and Marvel Rivals. Predict calibrated win rates for a draft, score many candidate next-picks in one batch, and recommend the best next pick. Render an interactive win-rate board for a draft. Read hero meta: tier lists, counters, synergies, head-to-head matchups, week-over-week momentum and ban rates. Read pro tournament data: most-picked and most-banned heroes, plus per-game box scores from parsed replays. Search batru.gg's live datasets and fetch the full document behind a result, including the model's public evaluation ledger.
When to use it
Useful when an assistant is asked about Dota 2, Deadlock or Marvel Rivals drafts, hero counters, synergies, tier lists or pro tournament results, and should answer with measured numbers rather than recalled guesses. The hosted endpoint is the recommended way to get the current tool set.
Requirements
Either the hosted remote endpoint at batru.gg, or the local PyPI package run with uv (Python 3.12+), which the README calls legacy and frozen at an early version. No authentication is declared. The optional BATRU_API_BASE environment variable defaults to the public site. Network access to batru.gg is required.
Before you install
The server is described as read-only and needs no credentials. The local package is outdated and exposes fewer tools than the hosted endpoint. The backend silently drops unrecognised hero names, so normalise names first to avoid misleading predictions. Tournament figures are small-sample records, not calibrated probabilities.

Installation

In SourceWeft

  1. Open Batru — Competitive Game Intelligence for AI Agents in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Use with an agent

Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.

Read https://sourceweft.com/mcp/gg.batru%2Fbatru-mcp/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.

Read the install guide the agent follows

Other MCP clients

Add this to your client's mcpServers config.

{
  "mcpServers": {
    "batru-competitive-game-intelligence-for": {
      "type": "http",
      "url": "https://batru.gg/mcp"
    }
  }
}

README

batru-mcp

[!IMPORTANT] Use the hosted endpoint — no install needed. Add https://batru.gg/mcp to Claude, ChatGPT, Cursor or any MCP client and you get the newest tool set (31 read-only tools across Dota 2, Deadlock and Marvel Rivals, always in sync with the live site). Setup guide: https://batru.gg/developers

The local PyPI package described below (uvx batru-mcp) is legacy and no longer updated — it is frozen at an early version with a reduced tool set. It still works (it proxies the same live API), but the hosted endpoint is the canonical distribution.

A minimal, read-only MCP server over batru.gg's live API, so your LLM can answer Dota 2 / Deadlock draft, counter and win-rate questions with real, calibrated model predictions instead of guessing from memory.

It is a thin wrapper around batru.gg's public endpoints — no model runs locally; every number comes from the same production model the website serves.

Why calibrated matters

batru.gg's model is trained on ~20M real matches and calibrated: a reported 60% win rate corresponds to an empirically observed ~60% win rate. We deliberately do not headline a raw "accuracy" number — accuracy alone is misleading for win prediction. What you get from these tools are probabilities you can trust at face value. The tool descriptions instruct the host LLM to report these numbers verbatim and never invent matchup data.

Tools

ToolWhat it does
lookup_hero(query, game="dota2")Normalise a name/alias/shortName to {id, displayName, shortName}. game ∈ {dota2, deadlock}.
predict_dota_winrate(my_heroes, enemy_heroes, my_side="radiant")Calibrated win-rate % for both teams (partial drafts OK; empty → 50/50).
recommend_dota_pick(my_heroes, enemy_heroes, my_side="radiant")Top-3 heroes to pick next, each with its calibrated win rate.
get_dota_counters(hero, limit=12)Real matchup table: who this hero beats / loses to, with win rate % and sample size.
predict_deadlock_draft(team0_heroes, team1_heroes)Calibrated win-rate % for a Deadlock 6v6 (6 heroes per team).

Hero names are accepted in any form (e.g. am, anti mage, Anti-Mage) and normalised internally — the backend silently drops names it doesn't recognise, so normalising first keeps predictions honest.

Install

Requires uv (or any way to run a Python 3.12+ package from PyPI):

bash
uvx batru-mcp           # fetches from PyPI and starts the stdio MCP server

Configuration is via the BATRU_API_BASE environment variable (default https://batru.gg) — you normally don't need to set anything.

Claude Desktop config

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

json
{  "mcpServers": {    "batru": {      "command": "uvx",      "args": ["batru-mcp"]    }  }}

Restart Claude Desktop; the batru tools appear in the tool picker. Claude Code: claude mcp add batru -- uvx batru-mcp.

Development

bash
git clone https://github.com/batrugg/batru-mcp && cd batru-mcpuv syncuv run batru-mcp        # run the server from the checkout (blocks, waiting on stdin)

For a Claude Desktop pointing at the checkout, use "command": "uv", "args": ["run", "--directory", "/absolute/path/to/batru-mcp", "batru-mcp"].

Tests

bash
uv run pytest            # offline: hero normalisation + draft assemblyuv run pytest -m live    # also hits the real batru.gg API

Prefer programmatic access from Python instead of MCP? pip install batru — the official batru SDK.

Source: README.md at commit 8d72267

Tools

17

lookup_hero

Look up a hero name

Normalise a hero name/alias/shortName to its canonical identity — or list the whole roster. Use this to turn messy user input ("am", "anti mage", "Anti-Mage") into the exact key batru.gg expects before calling the prediction tools. The backend SILENTLY DROPS hero names it doesn't recognise, so always normalise first. Not sure which heroes exist? Call it with an EMPTY query to get every hero of that game (displayName, shortName, aliases). Args: query: A hero name, alias, or short name. Empty string = list the full roster. game: "dota2" (default), "deadlock", or "marvel-rivals". Returns {id, displayName, shortName, game} for the best match, an error with `did_you_mean` candidates if nothing matches, or — for an empty query — {game, count, heroes:[{id, displayName, shortName, aliases}]}.

Parametersgamequeryrequiredcontextrequired

get_model_track_record

Model track record (public ledger)

Get batru.gg's public model evaluation ledger — every weekly test, wins AND losses. batru.gg publishes EVERY champion-vs-challenger evaluation of its Dota 2 model (promoted or rejected), auto-generated from the promotion gate's append-only log — nothing is hand-picked. Use this when a user asks whether the model is any good or how it is validated. HOW TO PRESENT IT: lead with CALIBRATION (ECE — lower is better; under 0.005 means a stated 60% wins ~60% of the time), then BCE. Do NOT headline raw accuracy on its own: Dota drafts are balanced by design, so pre-game draft-only accuracy sits in the high 50s for ANY model (always picking Radiant already scores ~53%) — calibration is the meaningful claim. Args: limit: Max evaluations to return, newest first (default 10). Returns {generated_at, ledger, total_evaluations, total_promotions, evaluations:[{decided_at, promote, challenger:{ece,bce,acc}, champion:{ece,bce,acc}, holdout_n}]}. Report numbers verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parameterslimitcontextrequired

analyze_draft

Analyze a game draft visually

Analyze a draft and render an interactive Batru win-rate board. Use this for visual draft, matchup, win-probability, or next-pick questions in Dota 2, Deadlock, or Marvel Rivals. It combines the existing prediction and recommendation tools into one stable structured result. Clients that support MCP Apps render a compact card; every other MCP client receives the same data as JSON/text. This tool is read-only. Dota 2 supports partial drafts and returns a calibrated probability plus calibrated next-pick win rates. Deadlock and Marvel Rivals only return a calibrated probability for a complete 6v6; partial drafts still return counter-pick recommendations labelled as empirical pair-data scores. Args: game: "dota2", "deadlock", or "marvel-rivals". my_heroes: Your team / Team 0 heroes (0-5 for Dota, 0-6 otherwise). enemy_heroes: Enemy team / Team 1 heroes (0-5 for Dota, 0-6 otherwise). my_side: Dota only — whether your team is Radiant or Dire. Returns the two rosters, prediction status, calibrated win rates when available, recommendations, metric labels, `source_url`, and `citation`. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredcontextrequiredmy_sidemy_heroesrequiredenemy_heroesrequired

predict_winrate

Predict draft win rate

Which team wins this draft — CALIBRATED win probability for Dota 2, Deadlock or Marvel Rivals. Backed by batru.gg's per-game production models (built on 140M+ real matches and calibrated: a reported 60% reflects a real ~60% empirical win rate — not a guess). Hero names are normalised internally. Per-game semantics: - dota2: full OR partial drafts (0-5 per side); empty draft = 50/50. `my_side` says whether my_heroes are Radiant or Dire. - deadlock / marvel-rivals: requires the full 6v6 rosters; `my_side` is ignored. For partial drafts use recommend_pick instead. Args: game: "dota2", "deadlock", or "marvel-rivals" (required). my_heroes: Your team's heroes (list of names; may be empty). enemy_heroes: The enemy team's heroes (list of names). my_side: Dota only — "radiant" (default) or "dire". Returns calibrated win-rate percentages for both teams. Report them verbatim; do not adjust them. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredcontextrequiredmy_sidemy_heroesrequiredenemy_heroesrequired

predict_winrate_batch

Score many candidate picks in one call

Score MANY candidate next-picks for a Dota 2 draft in ONE call — calibrated win rate + delta each. Use this instead of looping predict_winrate over candidates: the whole batch runs as a single model invoke server-side. For each candidate you get the CALIBRATED team win rate after adding it, plus the delta vs the draft without it (a full-roster sweep of up to 150 candidates fits in one call). Currently dota2-only. Args: my_heroes: Heroes your team already picked (0-4). enemy_heroes: Enemy heroes (0-5). candidates: Candidate heroes to evaluate (1-150). my_side: "radiant" (default) or "dire". game: Only "dota2" today. Returns {baseline_win_rate_pct, candidates:[{hero, shortName, win_rate_pct, delta_pp}]} sorted best-first. Report verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamecontextrequiredmy_sidemy_heroesrequiredcandidatesrequiredenemy_heroesrequired

recommend_pick

Recommend the next pick

Best heroes to pick next against this draft — Dota 2, Deadlock or Marvel Rivals. Per-game semantics: - dota2: top 3 picks scored by the CALIBRATED draft model (each with the calibrated team win rate after adding that hero); `my_side` applies. - deadlock / marvel-rivals: ranked by REAL observed pair data (win rate vs the enemy heroes, blended with duo rate alongside your picks) — an empirical heuristic, not the calibrated model; verify a full 6v6 with predict_winrate. Args: game: "dota2", "deadlock", or "marvel-rivals". enemy_heroes: Enemy heroes to pick against. my_heroes: Heroes your team already has. my_side: Dota only — "radiant" (default) or "dire". limit: Max recommendations (Deadlock/Marvel; Dota returns top 3). Report win rates verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredlimitcontextrequiredmy_sidemy_heroesenemy_heroesrequired

get_tier_list

Current-patch tier list

Best heroes right now — the current-patch tier list for Dota 2, Deadlock or Marvel Rivals. Empirical rankings from batru.gg's match aggregation over real games (not a guess); heroes come sorted best-first. Optionally filter to a skill bracket with `rank`. Args: game: "dota2", "deadlock", or "marvel-rivals". rank: Optional skill bracket. dota2: herald, guardian, crusader, archon, legend, ancient (and bands). deadlock/marvel-rivals: low, mid, high, or top (rank words like "oracle" or "eternus" also resolve). Empty = all ranks combined. limit: Max heroes to return (default 20). `total_matches` is the sample behind the ranking — cite it when stating how strong the signal is. Report numbers verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredranklimitcontextrequired

get_counters

Hero counters

Who counters a hero (and who it beats) — measured matchup win rates for Dota 2, Deadlock or Marvel Rivals. Opponents this hero performs BEST and WORST against, by real observed win rate with sample sizes — empirical meta data, not a guess. The hero name is normalised internally. Args: game: "dota2", "deadlock", or "marvel-rivals". hero: The hero to look up (name/alias/shortName). limit: Max matchups per direction (default 12). winrate_pct > 50 means `hero` beats that opponent. Report verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredherorequiredlimitcontextrequired

get_synergy

Hero synergies / duos

Best teammates / duo partners for a hero — measured together-win-rates for Dota 2, Deadlock or Marvel Rivals. Allies this hero wins MOST and LEAST with, by real observed win rate when picked together (with sample sizes). Covers Marvel Rivals team-ups too. The hero name is normalised internally. Args: game: "dota2", "deadlock", or "marvel-rivals". hero: The hero to look up (name/alias/shortName). limit: Max partners per direction (default 12). Report numbers verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredherorequiredlimitcontextrequired

get_matchup

Head-to-head matchup

One hero vs one opponent — the measured head-to-head win rate in Dota 2, Deadlock or Marvel Rivals. The direct matchup number ("does X beat Y?") from real games, with the sample size. Names are normalised internally. Args: game: "dota2", "deadlock", or "marvel-rivals". hero: Your hero. opponent: The enemy hero. Report the number verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredherorequiredcontextrequiredopponentrequired

get_hero_momentum

Hero momentum (rising / falling)

Which heroes are rising or falling right now — week-over-week win-rate movement. Momentum from batru.gg's rolling match aggregation for Dota 2, Deadlock or Marvel Rivals: the biggest movers of the current window, or one hero's trend when `hero` is given. Args: game: "dota2", "deadlock", or "marvel-rivals". hero: Optional — one hero's momentum instead of the movers list. limit: Max movers to return (default 10). Report numbers verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredherolimitcontextrequired

get_ban_rates

Ban rates

Most-banned heroes — measured draft-phase ban rates for Deadlock or Marvel Rivals. Args: game: "deadlock" or "marvel-rivals" (ranked Dota has no ban-phase export — for pro Dota bans use get_tournament_stats). limit: Max heroes to return (default 15). Report numbers verbatim. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgamerequiredlimitcontextrequired

get_tournament_stats

Pro tournament stats

TI / EWC pro tournament stats — actual most-picked, most-banned and win rates per hero. Real recorded results from the official drafts of every game at a covered pro Dota 2 event (The International, Esports World Cup, ...) — small-sample records, not calibrated probabilities. Answers "most picked / banned / contested at TI", "who won", "group standings". Args: event: Event slug, e.g. "ti-2026" (default), "ti-2025", "esports-world-cup-2026". limit: Max heroes per table (default 15). game: Only "dota2" today. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgameeventlimitcontextrequired

get_tournament_game_stats

Pro tournament box scores

Per-game BOX SCORES from parsed replays of a pro Dota 2 tournament. For each game: lane winners at 10 minutes, every player's hero, lane, K/D/A, GPM, net worth and key item timings, first blood, Roshans, and comeback flags. Filter by team, stage, or a single match_id. Args: event: Event slug (default "ti-2026"). team: Optional team-name filter. stage: Optional stage filter. match_id: Optional single-match filter. limit: Max games to return (default 3). game: Only "dota2" today. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersgameteameventlimitstagecontextrequiredmatch_id

fetch

Fetch a Batru document by id

Fetch the full live document behind a `search` result id. Ids look like `dota2:counters:axe`, `deadlock:tier-list`, `marvel-rivals:synergy:mantis`, `deadlock:ban-rates`, `dota2:tournament:ti-2026` or `model:track-record`. Returns {id, title, text, url} where `text` is the document's data as JSON (measured win rates / ranks with sample sizes). Report the numbers verbatim and cite `url`. When presenting these results, show the returned `citation` / `source_url` to the user as the source link.

Parametersidrequiredcontextrequired

get_more_tools

Get More Tools

Check for additional tools whenever your task might benefit from specialized capabilities - even if existing tools could work as a fallback.

Parameterscontextrequired

Version history

1
  1. v0.7.0LatestSep 16, 2026

Score

100% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 daysRequired25 pts
  • Auth declaredRequired10 pts
  • Tools listedRequired15 pts
  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts