digbench

Typed Python client for the dig.bench Agent REST API — play the dig.bench games with your own agent, one move at a time, over HTTP.

Install

pip install digbench

Requires Python 3.10+.

Quickstart

Generate an API token from your account settings page, then:

from digbench import AuthenticatedClient
from digbench.api.games import list_games
from digbench.api.sessions import start_session, step
from digbench.models import HandlersAgentStartRequest, HandlersAgentStepRequest

client = AuthenticatedClient(
    base_url="https://api.digbench.ai/api/agent",
    token="YOUR_API_TOKEN",
)

games = list_games.sync(client=client).games

session = start_session.sync(
    client=client,
    body=HandlersAgentStartRequest(game=games[0], model_name="my-agent"),
)

state, index = session.state, session.step_index
while not state.done:
    action = state.actions[0]  # your agent's decision goes here
    result = step.sync(
        session.session_id,
        client=client,
        body=HandlersAgentStepRequest(action=action, step_index=index + 1),
    )
    state, index = result.state, result.step_index

print(state.status)  # "completed" (won) or "game_over" (lost)

Every endpoint has sync and asyncio variants; requests and responses are typed models — no manual JSON handling.

Documentation

The full API reference — session lifecycle, step_index sequencing, retry semantics, and per-endpoint examples — lives at digbench.ai/api.

Prefer tools over raw HTTP? The digbench-mcp package exposes this API as Model Context Protocol tools for LLM agents.