mlbb-mcp

An MCP server for Mobile Legends: Bang Bang hero data, backed by ridwaanhall/api-mobilelegends. Built as a learning project for personal use, covering MCP server design, LLM tool design, grounded generation with citations, resilient API caching, and LLM-as-judge evals.

100% written by Claude; 100% reviewed by me.

Tools

Every stats-returning tool includes a citation block: source, retrieved_at, data_freshness, time_window_days, rank_tier. data_freshness is "fresh" or "stale" — the server serves cached data when upstream is unavailable rather than failing.

Setup

git clone <repo>
cd mlbb-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Copy .env.example to .env and fill in your keys (only needed for evals, not for the MCP server or CLI):

cp .env.example .env
# ANTHROPIC_API_KEY — required for both eval scripts
# OPENAI_API_KEY    — required for evals/comprehensive_evals.py (GPT judge) only

Run

CLI (no MCP server needed, useful for testing):

python cli.py heroes                  # list all heroes
python cli.py resolve lancelot        # resolve a name or ID
python cli.py resolve 47              # by numeric ID

MCP server (for Claude Desktop):

python server.py

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

{
  "mcpServers": {
    "mlbb": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["/absolute/path/to/server.py"]
    }
  }
}

Then restart Claude Desktop.

Tests

Unit and integration tests (no network, no API key):

pytest tests/ -v

The stale-cache tests cover every failure mode — network error, timeout, HTTP 5xx, non-JSON 200 (CDN error page), error envelopes, 4xx behaviour, cache poison guard, and recovery.

Evals

LLM behavior evals — separate from unit tests. They make real API calls and cost money, so run them intentionally rather than in CI. Both require ANTHROPIC_API_KEY (and OPENAI_API_KEY for the comprehensive suite) in .env.

evals/run_evals.py — focused behavioural checks

Two targeted evals that pre-inject tool results and grade the final response:

python evals/run_evals.py              # both
python evals/run_evals.py citation     # citation correctness only
python evals/run_evals.py fabrication  # fabrication refusal only

• citation_correctness — injects a real live tool result; checks Claude's response includes rank tier, time window, percentage, and source.

• fabrication_refusal — injects a ToolError; checks Claude refuses to invent stats rather than making something up.

evals/comprehensive_evals.py — full tool-selection + quality suite

33 questions across all 8 tools. Claude actually calls tools with real parameters; GPT grades each trace on tool selection, citation quality, and fabrication.

python evals/comprehensive_evals.py              # all 33, Haiku answerer (~$0.10, ~2 min)
python evals/comprehensive_evals.py --sonnet     # all 33, Sonnet answerer (~$1–2, ~5 min)
python evals/comprehensive_evals.py 5            # first 5 only (smoke test)
python evals/comprehensive_evals.py 5 build      # first 5, filter by category

Use Haiku by default during development; switch to --sonnet when you want to measure Sonnet's behaviour specifically. The judge is always gpt-4o-mini — the grading task is mechanical enough that a frontier judge model isn't needed.

Attribution

Game data © Moonton. API by ridwaanhall, hosted at mlbb.rone.dev.