Actual Play Database

An index of actual-play TTRPG credits organised by person.

Existing AP discovery tools index by show, game and channel. None index by performer. This one does, and it is built around the two things that omission breaks:

1. The filmography gap. Pull up a performer and see every credit across every show, season and episode — including one-off guest spots on small indie campaigns, not just the marquee work.
2. The alias gap. Performers change names. Other aggregators file the two names as unrelated people, which makes pre-name-change indie credits structurally invisible. Here, searching either name resolves to one person, and every credit records the name they were billed under at the time.

How it works

There is no database, no CMS and no auth backend. The data is flat YAML in git.

• Contribution is a pull request. Review is moderation, merge history is provenance.
• One file per entity — data/people/<slug>.yml, data/shows/<slug>.yml, and so on. Never a monolith: sharded files keep PRs atomic and diffs readable.
• IDs are human-readable slugs used as foreign keys, so the file path, the URL and the reference are all the same string.
• Credits live in the person file. That makes the filmography the literal source of truth, and makes the most common contribution — "I saw X in Y" — a single-file edit. Show cast lists are derived by inverting credits at build time.

Pages are statically generated, one real route per person and per show, because the discovery premise dies without individually indexable pages.

The CI gate

Flat files give you no referential integrity for free — a typo'd show id is a broken link that nothing catches. npm run validate is the replacement for every constraint a database would have enforced, and it runs on every PR:

• Schema validation on every file: types, required fields, enum values.
• Referential integrity: every show / season / channel / game / alias id referenced by a credit resolves to something that exists.
• ID uniqueness and slug format, and filename-must-equal-id.
• Duplicate-person heuristic: the same name on two people with no alias linking them gets flagged for maintainer review (advisory, not blocking).

npm run build runs the validator first and refuses to render a site from invalid data, so a broken reference can never reach the published pages.

Commands

npm install
npm run validate                              # the CI gate
npm run dev                                   # local site
npm run build                                 # validate, build, index for search
npm run filmography -- "Aabria Lipscomb"      # print a filmography to the terminal
npm run audit                                 # spot-check imports for bad data
npm run discover:person -- --person aabria-iyengar   # what are we missing for one person?
npm run report:discovery -- --person out/sweep.json  # fold a sweep into one report
npm run collect -- --sources                  # import sources + licence status

Data model

Entity	Lives in	Notes
Person	data/people/	Canonical identity. Holds its aliases and credits inline.
Alias	inside Person	name, alias_type, optional active range. The alias-gap fix.
Credit	inside Person	The polymorphic join. References a show, optionally a season and a free-text episode locator.
Show	data/shows/	Series under one or more channels. Holds its seasons.
Season	inside Show	ordinal, game reference, date range.
Channel	data/channels/	Producing entity: network, indie studio or solo creator — all first class.
Game	data/games/	edition is required; PF2e vs PF1e is a real discovery filter.

Episodes are not entities. A credit references an episode by a free-text locator ("S2E14") plus the show and season ids. This is what lets a single indie guest appearance surface on a performer's page without anyone having to catalogue an entire show episode by episode.

Main-cast credits reference the season. One-off guest credits additionally carry an episode locator — that distinction is the long-tail feature.

Provenance

Every credit carries a source declaring its tier, ranked by how close the source was to the thing happening: official → recording → participant → firsthand → reference → community. The tier renders next to the credit, and CI enforces the tiers that make claims about their own evidence: recording must carry a locator, firsthand must name who attested it.

There is deliberately no inferred tier, and no inference status anywhere. Nobody files a credit without having observed something — streams are watched — so inference never explains how a credit came to exist. And if a value was derived from sources, those sources are the account, so cite them; if it was derived from nothing, it isn't knowledge and shouldn't be a value. Fields nobody established are simply omitted: alias is optional, and such credits render as "billed name not established" rather than carrying a guess.

Note that firsthand — someone who watched the episode — outranks published reference works. It is closer to the fact. Wikipedia's filmography table lists Aabria Iyengar as a player on Pirates of Salt Bay when she ran it; anyone who watched would have got that right.

Corroboration is derived, never asserted. Each credit carries a list of sources and its status falls out of them: corroborated when two or more independent sources agree, single-source otherwise. Independence is checked by who attested a source or which page it cites, so one person filing twice does not count.

That list is what makes community verification work. A second listener who watched the same stream and files their own account moves a credit from single-source to corroborated — with no publisher involved anywhere.

The sourcing model is adapted from Wikipedia's, with one deliberate divergence: unpublished first-hand accounts are admissible here. Wikipedia's No Original Research rule assumes published sources reliably exist, which is false for the indie long tail — applying it would delete exactly the data no other index has. Testimony is admitted, labelled and ranked, never disguised as something stronger. The full reasoning is in POLICY.md.

Git history covers the rest.

Machine-readable access

The site is built to be fetched, not just crawled — an agent or script shouldn't need to parse rendered HTML to get the underlying facts.

• /llms.txt is the entry point: what this is, what's indexed right now, and links to everything below. Generated at build time so its counts never drift from what's actually in data/.
• /api/index.json is the machine-readable equivalent — a manifest of every endpoint, current counts, and how ids and season ordinals resolve.
• /api/people.json, /api/shows.json, /api/channels.json, /api/games.json are the full collections, same shape as the YAML, generated straight from the validated dataset.
• /people/{id}.json and /shows/{id}.json mirror each HTML page — fetch one record without downloading the whole collection.
• AGENTS.md is for AI agents specifically, not humans operating one. It covers how to file a contribution autonomously and, more importantly, which source tier an agent is and isn't allowed to claim — an agent that read a wiki page is a reference or community source, never firsthand, no matter how confident the page sounded.
• A sitemap and a robots.txt that explicitly welcomes AI crawlers (GPTBot, ClaudeBot, PerplexityBot, CCBot and others by name, not just a wildcard) round it out — the data is meant to be found, cited and corrected, not walled off from the tools that would do that at scale.

Importing

npm run collect has two modes.

Fan wikis (--wiki) are where the credits actually are. Fandom and Miraheze expose a real MediaWiki API with template-structured infoboxes, so this reads data rather than scraping prose:

npm run collect -- --wiki dimension20.fandom.com --page "Fantasy High" --show dimension-20 --season 1
npm run collect -- --wiki criticalrole.fandom.com --page "Campaign 1" --show critical-role --season 1

Crucially, these wikis mark guest players as their own infobox field, so one-off guest spots arrive already distinguished from main cast — the long tail, machine-readable. Field names differ per wiki (players/guest_players on Dimension 20, starring/sguests on Critical Role), so the field→role mapping is data in src/lib/sources/mediawiki.ts; add a row for a new wiki.

Cast fields interleave performers and characters, and every wiki formats that differently — so the parser doesn't read formatting at all. It resolves each linked page against the wiki's own categories (Cast / Voice Actors vs Characters vs Companies) in one batched API call, then walks the links in order: a person opens an entry, a character attaches to the entry before it, an organisation is dropped.

That one rule handles every layout encountered without knowing anything about any of them, and it is what stops a character being filed as a performer or a production company being filed as a person.

Show discovery (--discover) builds the show catalogue. A wiki already knows which of its pages are campaigns, so ask it — this enumerates pages transcluding the campaign infobox, no page names typed by hand:

npm run collect -- --wiki dimension20.fandom.com --discover --dry-run
npm run collect -- --wiki dimension20.fandom.com --discover --channel dropout

That finds 30 campaigns on the Dimension 20 wiki and 61 on the Critical Role wiki. Staged shows carry deliberately invalid placeholder ids for channel and game, so CI blocks a merge until a curator sets real ones — mapping a system string onto a game id needs a judgement about edition, which is required here.

YouTube (--youtube) reads a production's own video descriptions, which for shows with no fan wiki is often the only place a cast is written down — and is a better source than a wiki, because it's the producer describing their own episode. Credits from here are official tier.

export YOUTUBE_API_KEY=...   # free: console.cloud.google.com, YouTube Data API v3
npm run collect -- --youtube --playlist <PLAYLIST_ID> --show <show-id> --season 1 --dry-run

It reads only lines that name a role (DM:, Starring, Cast:, Produced by) and understands both X as Y and X (Y). A description with no cast line yields nothing rather than a guess. Episode credits roll up to the season — an index of who was in a show doesn't need a row per episode.

There is no keyless fallback: YouTube blocks unauthenticated access, and scraping around that would breach their terms.

Podcasts (src/lib/sources/podcast.ts) — Apple's lookup/search endpoints are public and keyless, and hand back the show's real RSS feed. Show metadata from there is official tier, because it's the production's own feed. It's also the route in for Spotify-only shows: read the identity from the page's og: tags, then look the title up on Apple.

Feeds carry title, description, author and episode count reliably. They rarely carry a structured cast, so hosts are read only from an explicit "hosted by" phrase and per-episode casts are not guessed at.

Wikidata (--qid, --search) is the identity backbone: stable QIDs, canonical names and, usefully, birth names. It carries almost no AP credits, so it is not used for them — but its works claims (P800, and a reverse P161 lookup) are read by the person-first sweep below, because a thin record of a show missing from the catalogue entirely is still the only pointer to it.

Wikipedia (npm run discover:person) is the only person-first source here, and it exists because every other adapter is show-first: they start from a show already in data/shows and read its cast down, so a show nobody has catalogued contributes nothing to anybody's filmography. That blind spot is worst for the people with the most credits, whose long lists look complete. Aabria Iyengar had 42 credits, a stored QID, a stored Wikipedia link and four missing series — all four named on the article this repo already linked to and had never read.

So this reads the person's article: filmography tables and prose both, since a credit lives in prose until someone builds a table row for it. Everything it finds is filtered through src/lib/sources/actual-play.ts, which classifies each candidate from the work's own article — only actual play belongs here, and a performer's filmography is mostly other things. Output is a review queue, never data: Wikipedia is reference tier and demonstrably fallible, which is a point this project makes with Wikipedia's own claim that Aabria played on a show she ran.

Internet Archive (npm run import:archive) is where shows go when they stop — productions whose sites are gone and whose feeds are dead. No other index has them, because there is no live source left to find. The search and metadata APIs are keyless.

Items are per-episode uploads rather than show records, so the work is grouping: creator is the producer, and a common title prefix across their uploads is a series. Systems come from the uploader's own subject tags; a series whose tags name no system is reported, not assigned one.

Sources that don't work

• dungeonsanddragons.com — serves an error page to non-browser clients; fully client-rendered with no structured data.
• dndbeyond.com — robots.txt disallows /api/, and the public pages carry no ld+json or embedded state. Their actual-play posts are prose.
• actualplay.world — no public API, and its sitemap lists only top-level pages, so there is no crawlable show index. Content renders client-side from React Server Component streams. Would need the operators' cooperation.
• Podchaser — has real creator credits, but the API needs a key issued under its terms; an adapter can't be built or tested without one.
• Fandom's cross-wiki search API — returns 403; wiki hosts have to be named explicitly rather than discovered.

Imports land in data/_incoming/ for review and become a PR like any other contribution. Curated data always wins: an import never overwrites an existing bio, links, credits or alias notes.

Sources that are not licence-cleared are refused by the collector until someone deliberately clears them in src/lib/sources/registry.ts.

Contributing

You do not need git, or YAML. The site has a /submit page: pick the person, show and season from what's already indexed, say how you know, and it writes the data file for you and deep-links a pre-filled GitHub issue. If you saw a different name on screen, it also writes the alias stanza.

Failing that, the add a credit and name change issue forms work too.

If you do use git, edit the YAML directly, run npm run validate, and open a PR. See CONTRIBUTING.md.

Stack

Astro, static output, deployed to GitHub Pages. The data layer (src/lib/) is deliberately framework-agnostic — plain TypeScript, Zod and YAML, with no Astro imports — so validation and the entity model survive a move to Next.js or anything else. Only the rendering layer is Astro-specific.