Distill MCP – Claude Code MCP를 통해 읽기 대기열을 팟캐스트로 전환

A Claude Code MCP server that turns saved articles into AI-generated podcasts, tailored to your interests. Save articles throughout the day. Generate a two-host conversational podcast on demand. MP3 lands in ~/Downloads , a sound plays, queue clears. Everything is local. git clone ~/code/distill cd ~/code/distill npm install Add Distill as an MCP server. The simplest way — from the directory where you want to keep your queue (your "reading project"): claude mcp add distill -- node /absolute/path/to/distill/index.js Or manually add to your ~/.claude.json / project .mcp.json : { "mcpServers": { "distill": { "command": "node", "args": ["/absolute/path/to/distill/index.js"] } } } Distill reads and writes files in the current working directory where Claude Code launched, so your queue and transcripts live alongside whatever project you're in. If you want a single global queue, set DISTILL_PROJECT_ROOT : { "mcpServers": { "distill": { "command": "node", "args": ["/absolute/path/to/distill/index.js"], "env": { "DISTILL_PROJECT_ROOT": "/Users/you/distill-home" } } } } Restart Claude Code. Confirm with /mcp — you should see distill connected. Easy mode: run /distill:setup inside Claude Code. It will walk you through interests, show style, TTS provider, and API key one question at a time. Manual mode: on first run Distill writes a default .distill file in the project root. Edit it directly: INTERESTS=entrepreneurship, cybersecurity, AI, startups SHOW_STYLE=The Briefing TTS_PROVIDER=openai OPENAI_API_KEY=sk-... OUTPUT_DIR=~/Downloads See .distill.example for a full annotated example. The spec describes .distill as both a file and a directory. Since that's not possible, Distill uses: .distill — config file (dotenv format).distill-data/ — data directory (queue.json ,transcripts/ ) Claude Code exposes MCP prompts with the prefix /mcp____ , so all Distill commands start with /mcp__distill__ : | Command | What it does | |---|---| /mcp__distill__setup | Interactive first-time configuration (interests, style, TTS, API key) | /mcp__distill__save | Fetch and queue an article | /mcp__distill__queue-list | Show the current queue | /mcp__distill__queue-clear | Empty the queue (no podcast) | /mcp__distill__podcast | Generate a podcast from a single article | /mcp__distill__digest | Generate a podcast from the queue, then clear | /mcp__distill__interests-add | Add an interest | /mcp__distill__interests-remove | Remove an interest | /mcp__distill__interests-list | Show current interests | Tip: you can usually type the first few letters (e.g. /distill-pod ) and let autocomplete fill the rest. All commands also work as natural language: "Save this to my podcast queue: https://…" "Make a Deep Dive from my queue" "Skip my usual interests, focus on the technical angle" | Style | Host dynamic | Vibe | |---|---|---| | The Briefing | Anchor + correspondent | Clean, newscast | | The Debate | Skeptic vs optimist | Punchy, opinionated | | The Deep Dive | Expert + curious learner | Educational | | The Watercooler | Two colleagues chatting | Casual | Override per session: "Make it a Debate style". | Provider | Notes | |---|---| openai | Default recommendation. Requires OPENAI_API_KEY . | elevenlabs | Most natural voices. Requires ELEVENLABS_API_KEY . | inworld | Character-focused. Requires INWORLD_API_KEY . | system | macOS say . Free, robotic. Requires ffmpeg (or sox ) to produce MP3: brew install ffmpeg . | - Article fetch: @mozilla/readability +jsdom — strips nav/ads/footers and returns clean text. - Script generation: the MCP prompt tells Claude Code's own LLM to read the queue, load the show style, and write the dialogue script inline. No separate Claude API key needed. - TTS: per-turn synthesis (one API call per dialogue turn, so each host gets its own voice), then concatenated into a single MP3. - Transcript: plain-text copy saved to .distill-data/transcripts/ for debugging. - Notification: macOS afplay plays a short system sound when the MP3 is ready. Distill is a local, stdio-based MCP server. Everything runs on the user's machine. There is no hosted backend. ┌───────────────┐ stdio ┌───────────────┐ HTTPS ┌──────────┐ │ Claude Code │◀──────────▶│ distill MCP │────────────▶│ TTS API │ │ (your LLM) │ │ (node process)│ │ (yours) │ └───────────────┘ └───────────────┘ └──────────┘ │ │ │ ▼ │ .distill / .distill-data/ ▼ writes script, calls tools Who pays for what: | Cost | Paid by | |---|---| | LLM that writes the dialogue | The user — it runs inside their Claude Code session, on their Claude subscription / API billing | | TTS audio generation | The user — they put their own OPENAI_API_KEY / ELEVENLABS_API_KEY / INWORLD_API_KEY in .distill | | Hosting | Nobody — stdio MCP servers don't need a server | This means anyone forking or installing Distill just clones the repo and points Claude Code at it. No account, no signup, nothing to deploy. - Claude Code only — not a standalone CLI - English only - macOS assumed for system sound - No JS-rendered or paywalled article support - No scheduling — generate on demand /mcp shows distill as failed — check the absolute path toindex.js in your MCP config and thatnpm install ran in the distill dir.- "Could not extract readable content" — the page is probably paywalled or JS-rendered. No retry in v1. - TTS fails with "missing API key" — uncomment exactly one TTS_PROVIDER= line in.distill and make sure its matching*_API_KEY is set. system TTS fails with "requires ffmpeg or sox" —brew install ffmpeg .