npr-cds-mcp

Ask NPR's Content Distribution Service (CDS) questions in plain words, from Claude or any MCP client. For NPR member stations. Not affiliated with or endorsed by NPR.

Source, docs, and the full guide for radio professionals: https://github.com/tmoody1973/npr-cds-openapi

Setup

You need a CDS token from NPR Member Partnership and Node.js 20 or newer.

npx -y npr-cds-mcp setup                              # asks for your token, then "which station are you?"
claude mcp add npr-cds -s user -- npx -y npr-cds-mcp  # Claude Code
codex mcp add npr-cds -- npx -y npr-cds-mcp           # Codex: the CLI, the IDE extension, and the ChatGPT desktop app

ChatGPT desktop app. It shares Codex's MCP settings, so the codex mcp add line above is enough. Without a terminal: in the app, open Settings, then MCP servers, then Add server; choose STDIO, name it npr-cds, command npx, arguments -y npr-cds-mcp. Run setup once first either way; every client reads the same saved token and station.

ChatGPT on the web (chatgpt.com) can't use it. ChatGPT's web connectors reach only MCP servers hosted at a web address, and npr-cds-mcp runs on your own computer. Use the desktop app, Codex, or Claude.

Any MCP client, as JSON:

{ "mcpServers": { "npr-cds": { "command": "npx", "args": ["-y", "npr-cds-mcp"] } } }

setup saves the token and your station under ~/.config/npr-cds/, readable only by you. npx -y npr-cds-mcp station changes the station later. NPR_CDS_TOKEN and NPR_CDS_HOME_STATION in the environment override the saved files. With a station set, other stations' content comes back marked display-only, per NPR's terms.

More: demo prompts for a newsroom.

What you can ask

• "What are the three newest Ladies First episodes?"
• "Latest NPR stories about AI"
• "Show me the newest KCRW stories"
• "Why isn't this story on our site?" (paste the url)
• "What labels does Radio Milwaukee actually use?"
• "What changed since yesterday?"
• "Newest This Bites episodes" (kind: podcasts)
• "NPR podcasts about AI" (each marked premium: play or link, never store)
• "Get me the latest NPR newscast"
• "/morning-prep" and "/newsletter-draft housing" (saved prompts)
• "What's WXPN's station id?"

Try it

1. "What's my home station?" → your station, no id typed.
2. "Three newest Ladies First episodes" → titles, dates, lengths, links, one call.
3. "Latest NPR newscast" → the hourly cut with a play link and a never-store note.
4. "Why isn't this story on the site?" + a url → labels, audio, image, teaser, problems in plain words.
5. "What labels does our station use?" → shows, podcasts, topics, tags with counts.
6. "What changed since 8am?" → new and updated, newest change first.
7. "NPR podcasts about AI" → episodes marked premium: play or link, never store.
8. "Read me that story" + a url → the paragraphs, clean; or the transcript; or an honest note.
9. "What did NPR cover on AI this week that we haven't?" → gaps, flagged for a human to judge.
10. Run the morning-prep prompt → a read-out-loud rundown with talk breaks.

Tools

| Tool | Give it | Get back | |---|---|---| | find_stories | words, a station, a show or podcast, a topic, a date window; kind: podcasts for episodes | compact hits newest first: title, teaser, date, link, audio, image (wide crop, resize template, credit), byline, collection names, rights note when premium or another station's | | read_story | a story url or CDS id | paragraphs in reading order, or the transcript, or a note that CDS has only teaser and audio; plus image, byline, owner, teaser, and premium flag | | check_story | a story url or CDS id | in CDS or not, labels by name, audio, image, teaser, problems in plain language | | station_labels | a station name | shows, podcasts, programs, topics, tags, categories it uses, with counts | | whats_new_since | a date or time, optionally a station | what was published or edited since, marked new or updated | | latest_newscast | nothing, short, or a station | newest newscast: time, length, stream link; never-store note when premium | | search_archive | words, a station, from and to dates | the archive in half-year windows, newest first | | coverage_scan | a topic and a window | what NPR and the network published, by station | | coverage_gap | a topic and a window | NPR's stories next to ours, marked localized or not | | network_pulse | a date window | which stations, shows, and topics were active, as counts with ids and last dates; never text | | find_station | a name, call letters, or city | station id, with city, state, band and music-only when the finder knows it | | find_collection | a topic, tag, show, or program name | collection id |

Four saved prompts chain the tools into workflows: morning-prep, weekly-prep (the week in review, what the network covered that we did not, the week ahead), show-prep (one show's episodes, the network on its beat, three questions for the host), and newsletter-draft. Plus one generated tool per CDS endpoint (queryDocuments, getDocument, …) for the full document when you need it.

Limits

CDS has no text search; word matching covers titles and teasers of the newest 300 to 1,800 stories per question, scoped to your home station and NPR unless you name a station. Names are learned as they pass through and cached under ~/.cache/npr-cds/. Audio is always a link, never a download. Content from other stations may be displayed with attribution and refreshed, not stored.

MIT.