Batru MCP server
{
"mcpServers": {
"batru": {
"command": "uvx",
"args": [
"batru-mcp"
]
}
}
}{
"mcpServers": {
"batru": {
"command": "uvx",
"args": [
"batru-mcp"
]
}
}
}{
"mcpServers": {
"batru": {
"command": "uvx",
"args": [
"batru-mcp"
]
}
}
}{
"servers": {
"batru": {
"command": "uvx",
"args": [
"batru-mcp"
]
}
}
}{
"mcpServers": {
"batru": {
"command": "uvx",
"args": [
"batru-mcp"
]
}
}
}file .mcp.json (in your project root)file claude_desktop_config.jsonfile ~/.cursor/mcp.jsonfile .vscode/mcp.jsonfile ~/.codeium/windsurf/mcp_config.jsoninstall uvx batru-mcpversion 0.2.1
What it does
An MCP server that exposes calibrated win-rate predictions, counter-pick data, and hero recommendations for Dota 2, Deadlock, and Marvel Rivals. It wraps batru.gg's live prediction API, allowing LLMs to answer draft and matchup questions with real model predictions rather than guessing.
- Get calibrated win-rate predictions for a Dota 2 draft given your team's heroes and the enemy team's heroes.
- Retrieve the top counter-picks against a specific hero, showing which heroes beat it and by what win-rate margin.
- Recommend the next hero to pick in a live draft based on both teams' current selections and win-rate impact.
Tools this server exposes
| Tool | Does |
|---|---|
| lookup_hero | Normalise a hero name, alias, or short name to a canonical {id, displayName, shortName} for Dota 2 or Deadlock. |
| predict_dota_winrate | Return calibrated win-rate percentages for both teams in a Dota 2 draft, accepting partial hero lists. |
| recommend_dota_pick | Suggest the top 3 heroes to pick next in a Dota 2 draft, ranked by their calibrated win rate given the current board state. |
| get_dota_counters | Return a matchup table showing which heroes beat or lose to a given hero, with win-rate percentages and sample sizes. |
| predict_deadlock_draft | Return calibrated win-rate percentages for a Deadlock 6v6 matchup given both teams' hero selections. |
Before you enable it
An MCP server runs with whatever access you give it. Start with a development account rather than production, and check which tools the server exposes before you let an agent call them unattended.