A personal Steam toolkit built to solve a specific problem: tracking year-to-date achievement progress against friends and generating AI-powered completion guides for games I’m chasing. Five CLI commands, a dual LLM backend, and enough Steam API work to access private libraries without touching privacy settings.
What it does
steam-achievements — Fetches your full friend list, pulls every played game per
player, and ranks everyone by achievements unlocked in the current calendar year.
steam-platinum — The main feature. For any game, it fetches your remaining
achievements, optionally grabs a community guide from Steam as grounding context, then
sends everything to an LLM — Claude or OpenAI, your choice — which returns a structured
strategy: categories, ordering, estimated hours, and tips. Results are persisted to
SQLite so previously generated guides are retrieved without re-calling the API.
steam-game, steam-friends, and steam-login — Browse achievements per
game, inspect your friends list, and authenticate via Steam.
The Steam auth problem
Steam’s API returns 401 for private libraries even with a valid developer key. Rather
than requiring users to make their profiles public, steam-login performs a full Steam
login flow (RSA key exchange, Valve Data Format parsing) and stores the resulting session
in the OS native keychain — Windows Credential Manager, macOS Keychain, or Linux Secret
Service. The session unlocks IPlayerService/GetOwnedGames for private libraries.
Achievement data still requires a developer key because
ISteamUserStats/GetPlayerAchievements doesn’t honor user access tokens.
Dual LLM provider
llm_provider.py defines an AsyncLLMProvider protocol with a single
complete(system, user) -> str method. Both the Anthropic and OpenAI implementations
satisfy it; the factory function reads LLM_PROVIDER from the environment and fails fast
if misconfigured:
LLM_PROVIDER=anthropic LLM_MODEL=claude-opus-4-5 steam-platinum --game 480
LLM_PROVIDER=openai LLM_MODEL=gpt-4o steam-platinum --game 480Concurrency model
Achievement fetching is async with 4 parallel requests per player by default, with automatic retry and exponential backoff on Steam’s 429 rate-limit responses.
Stack
Python 3.12+, httpx for async HTTP, SQLAlchemy + SQLite for persistence, Pydantic for
data contracts, keyring for cross-platform credential storage. Linted with ruff,
type-checked with pyright, tested with pytest.