SGU MCP
Allows semantic search using OpenAI's embedding API for higher quality query embeddings.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SGU MCP Search SGU transcripts for cold fusion."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
SGU MCP
An MCP server for The Skeptics' Guide to the Universe podcast — search transcripts, look up episodes, pull Science or Fiction, and search news items, all from inside Claude.
It ships in three forms:
A local MCP server (
npx sgu-mcp) for Claude Desktop / Claude Code / any MCP client.A remote MCP connector (Streamable HTTP) you can host so people connect their own Claude account and search the archive — no per-request cost to you.
A zero-backend web archive (
web/) — a static site where search runs entirely in the visitor's browser, with an optional "Ask Claude" panel (bring your own API key).
Requires Node ≥ 22.5 (uses the built-in node:sqlite — no native build step).
Quick start (local MCP server)
# Claude Code
claude mcp add sgu -- npx -y sgu-mcpOr add it to your MCP config manually (~/.claude/mcp.json or Claude Desktop's config):
{
"mcpServers": {
"sgu": {
"command": "npx",
"args": ["-y", "sgu-mcp"]
}
}
}Restart Claude, then try: "What was the Science or Fiction theme on SGU 1075?" or "Search SGU transcripts for cold fusion."
The live wiki/RSS tools work immediately. The fast local-archive tools need an index — download a prebuilt one in seconds:
npm run setup # downloads the prebuilt full-text index (data/sgu.db)…or build it yourself (see Building the archive).
Related MCP server: MCP Podcast Scraper
The two data layers
Local archive — every episode transcript scraped into
episodes/NNNN.md(YAML frontmatter + clean Markdown) and indexed into a SQLite FTS5 database (data/sgu.db) for instant, bm25-ranked, offline full-text search.Live wiki/RSS tools — for the newest episodes (whose transcripts aren't on the wiki yet) and as a fallback before the archive is built.
Data sources
No official SGU API exists (the website is a locked-down SPA). This server pulls from two reliable public sources:
Source | Used for |
sgutranscripts.org (MediaWiki API) | transcript search, episode segments, news items, Science or Fiction, full transcript text |
Podcast RSS feed (libsyn) | latest/recent episodes, release dates, audio URLs |
Note: Transcripts are volunteer-made and lag the feed by a few weeks. The newest episodes show up in
get_latest_episodes(RSS) before their transcript exists on the wiki.
Tools
Local archive (fast, offline, ranked — prefer these):
Tool | What it does |
| bm25-ranked FTS5 search over the whole archive (episode-level), highlighted snippets + metadata. Optional |
| Segment-level search — returns the exact moments with timestamp + speaker ("jump to the moment"). Filters: |
| Real occurrence count of a word/phrase across the archive, with breakdowns by year, by speaker, and top episodes. Answers "how many times did they say X?". |
| Natural-language / conceptual search. Blends vector similarity with BM25 (reciprocal-rank fusion). Needs the embedding index ( |
| Full Markdown doc (frontmatter + transcript) for an episode from the local archive. |
| Episode + segment counts, date range, and which embeddings are present. |
Live wiki / RSS (newest episodes + fallback):
Tool | What it does |
| Live full-text search of sgutranscripts.org. |
| One episode by number: title, date, rogues, guests, quote of the week, segment outline (timestamps), news items + links, Science or Fiction, audio URL. |
| The most recent episodes from the RSS feed (number, date, summary, audio). |
| The SoF theme + items + source links, and which item was the fiction (when machine-encoded; otherwise returns the segment transcript so the reveal can be read off). |
| Search the science news items the show has covered; returns topic, episode, link. |
| Cleaned transcript text — whole episode, or a single named section. |
Building the archive
If you'd rather build the index from scratch instead of npm run setup:
npm run fetch # scrape all ~1000 transcripts -> episodes/*.md (resumable; skips existing)
npm run index # build data/sgu.db (SQLite FTS5) from the .md corpusfetch is polite (limited concurrency, rate-limited, retries) and resumable — re-running only
fetches missing episodes. Use --force to re-scrape, or --only 1075,1074 for specific episodes.
Segment index & semantic search
npm run index builds two layers from the corpus: the episode FTS index and a
segment index (one row per speaker turn, ~178k rows) that powers search_segments
(timecoded, per-speaker) and count_mentions (true occurrence counts).
For conceptual / natural-language search, build the embedding index:
npm run embed # local model (default) — no API key, no cost
EMBED_PROVIDER=openai npm run embed # needs OPENAI_API_KEY (one-time, ~cents)
EMBED_PROVIDER=voyage npm run embed # needs VOYAGE_API_KEYEmbeddings are episode-level (one vector each), computed once. semantic_search then
blends them with keyword ranking. The provider for query embedding is set by
EMBED_PROVIDER (default local, via @xenova/transformers, an optional dependency).
A note on Science or Fiction answers
The fiction item is reliably structured only when the transcription bot encoded it
(answerKnown: true). Otherwise the reveal lives in the discussion prose, so the tool returns the
SoF segment transcript and answerKnown: false — Claude reads the answer from it.
Remote MCP connector (search with your own Claude account)
Host the server over Streamable HTTP and people can add it as a connector in Claude (Desktop / Code / Team / Enterprise), searching the archive with their own Claude subscription.
npm run build
SGU_MCP_TOKEN=$(openssl rand -hex 24) npm run start:http # serves POST /mcp on :8788Health check:
GET /healthzMCP endpoint:
POST /mcp(stateless Streamable HTTP — a fresh server per request)Auth: if
SGU_MCP_TOKENis set, clients must sendAuthorization: Bearer <token>. If it's not set, the server binds to loopback only, so you can't accidentally expose an unauthenticated endpoint.
Add it in an MCP client with a bearer header, e.g. Claude Code:
claude mcp add --transport http sgu https://your-host.example.com/mcp \
--header "Authorization: Bearer <token>"Deploy on Render: render.yaml defines the connector as a Node web service
(sgu-mcp-connector). It downloads the prebuilt index at build time (no scraping), serves the same
12 tools, and reads SGU_MCP_TOKEN from the dashboard.
Public Claude.ai connector? Anthropic's hosted Claude.ai expects a full OAuth 2.1 flow for custom remote connectors. The bearer-token mode here is perfect for self-hosting and for Desktop/Code/Team custom connectors; putting an OAuth proxy (or an MCP-aware gateway) in front is the next step for a public listing. The tool layer is unchanged either way.
The fan-facing web archive (web/)
A fully static, zero-backend search site. Search runs entirely in the visitor's browser — no server, no API key — so it costs nothing to run no matter how much traffic it gets.
How it works: the FTS5 database is served as a static file and queried in-browser via
sql.js-httpvfs(SQLite compiled to WASM). HTTP range requests mean the browser only downloads the few KB of DB pages each query touches — not the whole file.What fans can do:
Full-text search across every transcript, bm25-ranked, with highlighted snippets.
Counting questions — "how many times was homeopathy mentioned in 2024?" → a number, a per-year breakdown, and the episodes themselves.
Filter by year; jump straight to the transcript or the audio.
"Find by meaning" (semantic search) with two modes, so it works for everyone and never costs you per query:
Free — embeds the query in the visitor's browser via
@xenova/transformers(loads a ~25 MB model once). No key, no cost; a little slower.Best — the visitor pastes their own OpenAI key; their browser calls OpenAI for the query embedding (fast, highest quality). Requires that you published OpenAI doc vectors once (
EMBED_PROVIDER=openai npm run embed); otherwise the panel says so and Free mode still works. Doc vectors ship inside the static DB, so ranking happens entirely in the browser.
Optional "Ask Claude" — a BYOK panel where the visitor pastes their own Anthropic API key; their browser calls the API directly (never this site), and Claude answers from the top transcript excerpts, citing episode numbers. See SECURITY.md.
Build & run the web archive locally
npm run setup # download the prebuilt index (or fetch + index yourself)
npm run web # prepares web/sgu.db (single-file, vacuumed, range-ready)
npm run web:serve # http://localhost:8787 (range-aware static server)The web/vendor/ files (sql.js-httpvfs) are committed so the site is self-contained and works
offline — browsers block cross-origin Worker scripts, so they must be served same-origin.
Deploy the website to Render
render.yaml's sgu-archive static site downloads the prebuilt index and prepares the browser DB —
no scraping at deploy time, so deploys are fast and gentle on the volunteer wiki:
buildCommand: npm ci && npm run setup && npm run web
staticPublishPath: ./webHow the archive stays fresh
.github/workflows/publish-db.yml runs monthly (and on demand) in GitHub's cloud — works with
your Mac closed. It scrapes new transcripts, rebuilds the index, and publishes it to the db-latest
GitHub Release. Everything downstream (npm run setup, the connector, the website) just downloads
that artifact — one polite scrape feeds them all. If you set a RENDER_DEPLOY_HOOK_URL repo secret,
it also pings Render so the website redeploys with the new data.
Development
npm run dev— run the stdio server from source withtsx(no build step)npm run dev:http— run the HTTP connector from sourcenpm run smoke— live test against SGU sourcesSource:
src/—server.ts(the 12 tools),index.ts(stdio entry),http.ts(HTTP entry),wiki.ts(MediaWiki client),rss.ts(feed),parse.ts(wikitext parsers),db.ts(FTS5 + segments + vectors),segments.ts(speaker-turn parser),embeddings.ts(providers)
See CONTRIBUTING.md. Code is MIT (LICENSE); transcript/audio content belongs to their authors — this is an unofficial fan tool.
Available Tools
12 toolsarchive_statsSGU local archive statsA
Report how many episodes are in the local indexed archive and the date range covered. Use to check whether the archive is built and how complete it is.
Inputs: no named parameters.
count_mentionsCount SGU mentions of a termA
Count how many times a word or phrase is actually said across the whole archive — a real occurrence count, not just how many episodes match. Returns the total, segments/episodes matched, and breakdowns by year, by speaker, and the top episodes by frequency. Use for questions like 'how many times have they said homeopathy?' or 'who says "awesome" the most?'. Matches the stem and its inflections (e.g. homeopath → homeopathy, homeopathic). Covers transcribed episodes only (a few weeks behind release).
Inputs: term (required), top_episodes.
get_episodeGet an SGU episodeA
Get a structured overview of one episode by number: title, date, rogues present, guests, quote of the week, the segment outline (with timestamps), news items (with source links), Science or Fiction, audio URL, and links. Combines the RSS feed (recent metadata + audio) with the transcript wiki (segments + details). Use get_transcript for the full text.
Inputs: episode_number (required).
get_episode_markdownGet archived episode markdownA
Return the full Markdown document (YAML frontmatter + clean transcript) for an episode from the local archive. Frontmatter includes date, rogues, guests, theme, the Science-or-Fiction answer, news items with links, audio URL, and source. Use after search_episodes to read the full text.
Inputs: episode_number (required).
get_latest_episodesGet latest SGU episodesA
List the most recent episodes from the podcast RSS feed: number, title, date, summary, audio URL. Use to find the newest episode or recent ones. Note: transcripts lag a few weeks behind release.
Inputs: limit.
get_science_or_fictionGet Science or FictionA
Get the Science or Fiction segment for an episode: the theme, the items (with source links), and which item was the fiction WHEN it is machine-encoded (answerKnown=true). When answerKnown=false, the reveal isn't in structured data — the transcript of the segment is included so the answer can be read off the discussion.
Inputs: episode_number (required).
get_transcriptGet SGU transcript textA
Get the cleaned transcript text of an episode. By default returns the full transcript (can be long). Pass a 'section' substring (e.g. 'Science or Fiction', a news item title, or 'Intro') to return only that section's text. Use get_episode first to see the segment outline.
Inputs: episode_number (required), section.
search_episodesSearch the SGU archive (local index)A
Fast, bm25-ranked full-text search over the LOCAL archive of episode transcripts (the indexed .md corpus). Prefer this over search_transcripts for general questions — it's instant, offline, and ranked, and it returns highlighted snippets with episode metadata. Optionally restrict to a field: 'transcript' (default scope is all), 'news' (news-item titles), or 'title'. Falls back with a note if the index hasn't been built yet (run npm run setup or npm run fetch && npm run index).
Inputs: query (required), limit, field.
search_news_itemsSearch SGU news itemsA
Search the science news items the show has covered. Each news item is its own topic page tagged with the episode number. Returns topic title, episode number, and link. Use to find which episode covered a topic, or to survey coverage of a subject.
Inputs: query (required), limit.
search_segmentsSearch SGU segments (timecoded)A
Fine-grained full-text search over individual speaker turns in the local archive. Unlike search_episodes (which returns whole episodes), this returns the exact moments — each result has the episode, date, segment/section, the timestamp to jump to, the speaker, and a highlighted snippet. Use for 'find the moment when…', quoting who said what, or narrowing within an episode. Optional filters: episode, speaker (e.g. 'Steve', 'Cara'), year.
Inputs: query (required), limit, episode, speaker, +1 more.
search_transcriptsSearch SGU transcriptsA
Full-text search across all SGU episode transcripts and topic pages on sgutranscripts.org. Use for questions like 'every time they discussed CRISPR' or 'what did they say about cold fusion'. Returns matching pages with snippets, the episode number, and a wiki URL.
Inputs: query (required), limit.
semantic_searchSemantic search SGU episodesA
Concept-level search that finds episodes by meaning, not just keywords — ask in natural language (e.g. 'episodes about the ethics of de-extinction' or 'when they got frustrated with science denial'). Blends vector similarity with keyword (BM25) ranking via reciprocal-rank fusion. Returns episodes with title, date, theme, and a fused relevance score. Requires the embedding index (npm run embed); the provider is set by EMBED_PROVIDER (default 'local', no API key). Prefer search_episodes/search_segments for exact words; use this for fuzzy, conceptual questions.
Inputs: query (required), limit, mode.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 12 tools
Each tool has a clearly distinct purpose. Tools like get_episode, get_episode_markdown, and get_transcript serve different retrieval needs; search tools are differentiated by scope and method (full-episode vs. segment, keyword vs. semantic). Descriptions further clarify overlap.
All tool names follow a consistent verb_noun pattern using snake_case (e.g., get_episode, search_segments). The convention is uniform across all 12 tools, enhancing predictability.
12 tools is well-scoped for a podcast archive and search server. It covers browsing, detailed retrieval, multiple search strategies, and statistics without being bloated or sparse.
The tool set covers core operations: episode retrieval, transcript access, various searches, and mention counting. A minor gap is the lack of a direct 'list all episodes' tool, but search tools can compensate.
Maintenance
Related MCP Connectors
Augments MCP Server - A comprehensive framework documentation provider for Claude Code
An MCP server that provides tools to discover and retrieve podcast episodes transcripts.
Hosted MCP server for live public-data APIs and Skills for AI agents.
Bible MCP server for ChatGPT, Claude and Grok: Scripture, study notes, Strong's Greek and Hebrew.
Related MCP Servers
- AlicenseDqualityDmaintenanceMCP server that provides Claude AI assistants with the ability to search the web, get news, and perform research using the You.com API.42MIT
- AlicenseAqualityDmaintenanceAn MCP server that scrapes and transcribes podcast episodes from YouTube or RSS feeds using Deepgram's Nova-2 model. It allows users to track podcasts for new episodes, manage transcripts, and generate personalized summaries through Claude.101MIT
- FlicenseAqualityCmaintenanceMCP server that exposes Discord and Twitch/Chatty chat log query tools to Claude Code, enabling searching messages, channel stats, user messages, and events from chat logs.7-
- AlicenseNot gradedqualityDmaintenanceAn MCP server that lets Claude answer ARK: Survival Ascended / Evolved questions by querying the ARK community wiki for taming, crafting, spawns, and stats.MIT