Zum Inhalt

questlog is an AI-first game tracker with a built-in MCP endpoint, on Next.js and SQLite

Context and Problem Statement

Marcel plays long Soulslikes (Lies of P on the Switch 2, earlier Dark Souls, Elden Ring, Sekiro) and loses track of the threads he started: an NPC waiting for an item, a choice he still has to make, which weapon part he kept for later. Playing Lies of P he kept this in his Trilium LLM wiki: a quest log page, a build page and a mechanics page, written by Claude while he talked. It worked, but each update rewrote a whole page, needed a write lock, and "which quests are still open?" was a read of free text, not a query. He wants the same experience, talking to Claude while playing, backed by a tool built for it.

Decision Drivers

  • Capture is by talking to Claude (and sharing screenshots), never by forms.
  • Structured questions ("what is still open?", "how did my Technique grow?", "does this weapon fit my build?") must be answered reliably.
  • No spoilers: the tool must never show what the player hasn't found (values V1).
  • Works for any game, not just Lies of P (V8).
  • Each player's log is theirs alone, enforced by the server (V2).
  • One small thing to run in the homelab, with one data volume.

Considered Options

  • Keep it in Trilium. Already works and is linked with the rest of the wiki. But every change rewrites a page, quest status is free text, and the spoiler rule is a convention the assistant has to remember.
  • Add games to freundebuch. Same stack and sign-in, nothing new to run. But freundebuch is about the people in a household's life; a quest log would blur what it stands for.
  • An existing tracker (Notion templates, spreadsheet trackers, game-specific checklist sites). Built around checklists of everything in the game, which is the opposite of spoiler-free, and none is reachable by Claude in a structured way.
  • Build our own, AI-first, modelled on freundebuch. (chosen)

Decision Outcome

Chosen option: "Build our own, AI-first, modelled on freundebuch", because it keeps the way of working that already proved itself (talking while playing), turns progress into data, and makes the spoiler rule a property of the query layer.

  • Capture through the built-in MCP endpoint (POST /mcp), added as a custom connector in claude.ai.
  • Stack and structure from freundebuch: one Next.js app (standalone output, one Docker image), SQLite (better-sqlite3, kysely) on one volume, a framework-free backend/ behind three surfaces (MCP, REST /api/v1, pages), sign-in through OIDC (authentik) and its own OAuth server for claude.ai. next and the MCP SDK pinned exactly.
  • Patterns from freundebuch's ADRs: every tool result carries next (freundebuch ADR 0010); a play session is its own record holding the transcript, and everything filed from it points to it (ADR 0015); a shared screenshot is read, filed as text and deleted (ADR 0014).
  • Spoiler gate: every quest, item and note has a spoiler level; one function decides what a read returns, like freundebuch's visible(). "Did I miss anything?" is its own tool and needs an explicit spoiler opt-in.
  • Language: code, API, tool names and docs in English; the UI in English first, with room for other languages. Game terms keep their localized names next to the English ones.
  • One log per player, not per household (decided by Marcel, 2026-10-30): everyone who can sign in has their own games, playthroughs and state, and nobody sees another player's log. This differs from freundebuch, where a household shares one book. Every read is scoped to the signed-in player, in the same query layer as the spoiler gate.
  • Sign-in through any OIDC provider: authentik first, as in freundebuch, but configured as a generic OpenID Connect client so another provider can be used later without code changes.

Domain (first sketch)

Record Holds
Game Title, platform; the stat names and item kinds the game uses, as data
Playthrough One run of a game (NG, NG+), language of the game
Session The transcript of one play session; everything below points to the session it came from
Quest NPC or name, where it starts, status, dated log entries, a vague next step, spoiler level
Build The plan (main stat, targets) and dated stat snapshots
Item Weapon or part, where and when found, scaling, equipped or not, Claude's verdict with sources

MCP tools (first sketch)

  • record: file one session and split it into quest entries, stat snapshots, items and decisions.
  • briefing: "where was I?" for a playthrough: open quests, current build, the last session.
  • open_threads: quests that are started and waiting on the player.
  • item: look up an item the player has, or compare a new one with the build.
  • missed: what the player may have missed; only with an explicit spoiler opt-in.
  • lookup: fetch facts about an item or quest the player has from a source such as a wiki (ADR 0002).
  • correct: fix or remove anything filed.

Consequences

  • Good, because "what is open?" and "how did my stats grow?" become queries, not readings of a page.
  • Good, because spoiler safety is enforced by the server, not remembered by the assistant.
  • Good, because freundebuch's code and conventions give a working start.
  • Bad, because it is a second app to maintain next to freundebuch. Sign-in and OAuth come from Better Auth (ADR 0004); the rest of the plumbing is copied.
  • Bad, because what Claude finds on the web about items and quests may be wrong; questlog keeps the sources so the player can check.

Later

  • Import the existing Trilium pages (Lies of P quest log and build) as the first playthrough, once questlog can record them (decided by Marcel, 2026-10-30).