exyu-mcp

An MCP server that gives AI agents (Claude, ChatGPT, and others) the ex-Yugoslav cultural context they're currently missing.

The problem

LLMs are fluent in ex-YU languages but culturally thin. They know the five most-famous lines of a cult film, not the 200th-most-quoted one, the exact wording, or who actually said it. The exact quote, the correct attribution, the slang meaning, the modern usage — that long tail is missing or hallucinated. This repo builds the layer that fills it: a factual, citable, safe-by-construction cultural-reference database exposed as an MCP server, so any connected agent (Claude, ChatGPT, others) can look it up instead of guessing.

Related MCP server: BENZEMA's Personal Knowledge Card

Scope

This repo is deliberately narrow: cultural reference resolution only — quotes, attribution, timestamps, slang meaning, cultural weight. Safe, factual, sourced. Full design in brainstorming/.

We don't try to convince or configure the host platforms — we build the data and the MCP, ship it, and let each platform decide whether to use it. The lever we control is trust: transparent sourcing and per-field provenance, not permission-seeking.

Explicitly out of scope here: shaping how a model talks — tone, directness, swearing as normal register rather than something to sanitize. That's a different kind of problem (what a model is willing to say, not what it knows) and needs a different delivery mechanism than an MCP tool call to someone else's chat app. Planned as a separate future project — a skills/harness layer for a custom chat app or direct API use, potentially reusing this repo's MCP/data. Not started; revisit once this repo's MVP is live.

Current-phase decisions

• Copyright/licensing: deferred. We go bold on short-quote usage for now and treat it as a solved problem — there's no way around ingesting copyrighted subtitles/lyrics for the payload this project needs, and stalling on it blocks everything else. The full risk/tier analysis is kept in brainstorming/09-risks-and-licensing.md for when this needs a real answer (before any wide/public distribution) — it's deferred, not forgotten.

Status

Working MVP, not yet through live acceptance. The Phases 0–2 build from the design spec is implemented and green in CI:

• TypeScript MCP server exposing resolve_reference over Streamable-HTTP (src/server/http.ts) and an npx stdio twin (bin/exyu-mcp.ts).

• Postgres + pgvector/pg_trgm/tsvector storage, with a trigram → FTS → vector cascade and fully-cited results (work, speaker, timestamps, countersign, meaning, modern usage, per-field provenance).

• YAML records validated against a locked JSON Schema, ingested into Postgres transactionally and idempotently.

• Docker Compose deployment behind Caddy (TLS + rate limiting), with push-to-main CI/CD to a Hetzner box.

Not done yet: the live vertical-slice acceptance test — the Valter record is still a stub with ⚠ TO-VERIFY fields (countersign, speaker, timestamps) that need real subtitle data, and the 06-plan.md acceptance test (type "Vazduh gori ko da…" into a real Claude/ChatGPT chat and get the resolved record back) has not been run against a live deploy.

Full design: docs/superpowers/specs/ and docs/superpowers/plans/; original design notes in brainstorming/.

Quickstart

cp .env.example .env          # then fill in the values below
npm ci
docker compose up -d db       # or `docker compose up -d --build` for the full stack
npm run build                 # required: `npm run migrate` runs the compiled output
npm run migrate               # apply db/migrations/
npm run validate              # schema-check every records/**/*.yaml
npm run ingest -- records/film/ref_valter_vazduh_trepti.yaml
npm run dev                   # http://localhost:8787/mcp  (or `npm start` after build)

Other scripts: npm test (vitest; the DB-backed tests need Docker for testcontainers), npm run migrate:dev (migrate via tsx, no build needed).

Environment variables

The refs.embedding column is vector(1536), matching the default OpenAI text-embedding-3-small model. Switching EMBEDDING_PROVIDER to a different-dimension model (voyage-3 is 1024) requires altering that column; ingest and resolve both fail with an explicit dimension-mismatch error otherwise.