Best AI Agent Skills for Podcast Hosts in 2026
A. Frans
Published May 26, 2026
Table of Contents
- 01Quick comparison
- 02elevenlabs-mcp: voice generation that doesn't sound off
- 03deep-research: the show-prep skill that scales
- 04mcp-server-youtube-transcript: research at the source
- 05vexa: meeting recording that doesn't lock you to one platform
- 06claude-seo: show notes that earn search clicks
- 07A combined workflow
- 08Auditing skills before you install them
- 09FAQ
Most podcast hosts I know spend more time on the surrounding workflow than on the actual recording: guest research, transcript cleanup, show notes, social repurposing, sponsor pitch prep. Claude Code skills sit in a different category from the editing tools in the previous article. They run inside your terminal, they orchestrate other tools, and they handle the parts of the job that aren't quite "production" but eat your week anyway.
I've been running five of these skills as part of my own pre-record and post-record routine for the last two months. Here's what each one does, how to install it, and where the rough edges are.
Quick comparison
| Skill | What it does | GitHub source | License |
|---|---|---|---|
| elevenlabs-mcp | Generate voice, clone voices, build intros | github.com/elevenlabs/elevenlabs-mcp | MIT |
| deep-research | Multi-source guest and topic research | github.com/wuyongjia/deep-research | MIT |
| mcp-server-youtube-transcript | Pull transcripts from any YouTube video | github.com/anaisbetts/mcp-youtube | MIT |
| vexa | Record and transcribe Zoom/Meet/Teams calls | github.com/Vexa-ai/vexa | Apache 2.0 |
| claude-seo | Optimize show notes and episode pages for search | github.com/anthropic-skills/claude-seo | MIT |
elevenlabs-mcp: voice generation that doesn't sound off
This is the official MCP server from ElevenLabs and it gives Claude Code direct access to voice generation, voice cloning, and audio manipulation. For podcasters, three use cases matter:
- Generate an intro/outro voice that matches your show without re-recording
- Build promo trailers in your own voice (cloned with consent) without re-recording
- Generate sponsor read variations to A/B test which version converts
Install:
claude mcp add elevenlabs npx @elevenlabs/elevenlabs-mcp
You'll need an ElevenLabs API key. Free tier gives 10,000 characters per month, paid tiers start at $5/mo. The voice cloning quality has improved enough by 2026 that I'd stop trying to spot clones by ear. That's a problem for trust on the internet, but a useful capability for podcasters who already own their voice.
Security note: only clone your own voice. Cloning a guest's voice without explicit written consent runs afoul of state right-of-publicity laws in the US and similar laws elsewhere. ElevenLabs has consent-verification for cloning, but the legal risk sits with you.
deep-research: the show-prep skill that scales
If you do interviews, prep is the unglamorous time sink. Three hours reading the guest's past interviews, finding their best stories, surfacing what they haven't been asked yet. The deep-research skill runs an 8-phase pipeline (search, synthesize, fact-check, format) that gets you 80% of that prep in 15 minutes.
The way I use it: I drop the guest's name and a few keywords into Claude Code, the skill kicks off, and it returns a brief with:
- Guest's most-cited past quotes and which interviews they came from
- Topics they've already covered in detail (so I don't repeat)
- Open questions the guest has hinted at but hasn't answered fully
- Recent controversies, news, or context I should know
Install:
git clone https://github.com/wuyongjia/deep-research ~/.claude/skills/deep-research
Where it falls short: the skill is only as good as the public web. Guests with thin online presence (most non-celebrities) get sparse briefs. For those, the skill saves maybe 30 minutes; for high-profile guests, it saves three hours. The other miss is non-English source material. If your guest gave a great interview in Portuguese to a small Brazilian podcast, deep-research probably won't surface it. That's a gap if you book international guests regularly.
mcp-server-youtube-transcript: research at the source
This one solves a specific problem. You remember your guest said something interesting on a podcast 18 months ago and you want the exact quote. Open the YouTube video, paste the URL, get the full transcript with timestamps. No need to scrub through the video.
Install:
claude mcp add youtube-transcript npx @ngonjia/mcp-youtube
Combined with deep-research, the workflow is: research finds the relevant interview, transcript skill pulls the exact wording, you reference the quote in your pre-interview email or in your opening question. Guests notice when you've sat with their past work in detail.
Limitations: only works on YouTube videos with captions (auto-generated or manual). Live-only video, private uploads, and the rare upload that disables captions won't work.
vexa: meeting recording that doesn't lock you to one platform
Vexa is open-source meeting recording across Google Meet, Microsoft Teams, and Zoom. For podcasters doing pre-interviews (the 20-minute call where you and the guest agree on topics and vibe), this is invaluable. Record the pre-call, get a transcript, harvest the natural-sounding versions of stories the guest told you before the official recording.
Install:
# Self-host the API
git clone https://github.com/Vexa-ai/vexa
cd vexa && docker compose up -d
# Then add the MCP server
claude mcp add vexa npx @vexa-ai/mcp-server
The self-hosting requirement is real. Vexa runs as a Docker stack and you need a server (or beefy local machine) to run it. For podcasters who don't want to manage infrastructure, there's a hosted option in beta at vexa.ai with a free tier.
Consent matters. Recording a call without disclosure is illegal in two-party-consent states and unethical everywhere. Always tell the guest the pre-call is being recorded.
claude-seo: show notes that earn search clicks
Show notes are where most podcasters lose. Either you skip them and miss search traffic, or you write them and they sound like AI-generated junk that Google demotes. Claude-seo (Anthropic's universal SEO skill) handles the optimization layer: keyword research, on-page structure, schema markup, internal linking strategy.
Install:
claude plugin install claude-seo
For a podcast episode page, the workflow is: 1. Drop the transcript + episode title into Claude Code 2. Ask the skill to produce show notes optimized for a specific query 3. Skill outputs: H1 + H2 structure, intro paragraph, 5-bullet summary, quote pull-outs, schema markup, internal link suggestions
The output isn't publish-ready prose. You still rewrite it in your voice. But the structure and the keyword targeting come out of the skill, which is where most hosts get it wrong.
A combined workflow
Pre-record (Sunday afternoon):
- Run deep-research on each guest scheduled this week
- Pull transcripts of their best 2–3 past appearances with mcp-server-youtube-transcript
- Generate sponsor read variations with elevenlabs-mcp for A/B testing this week's episode
Day of recording:
- Use vexa to capture the pre-call (with consent)
- Record the main episode in your DAW of choice
Post-record (Tuesday):
- Generate show notes with claude-seo based on the transcript
- Run elevenlabs-mcp to produce promo audio clips in your own voice
- Edit episode in tools you already use (Descript, Adobe Podcast; see our tools roundup)
Total skill cost: ~$10/mo (ElevenLabs free tier covers most use cases; everything else is free open source). Time saved per episode: 4–6 hours.
Auditing skills before you install them
The five above are reasonably safe. Three come from named companies (Anthropic, ElevenLabs, Vexa team), the other two are popular community skills with active maintainers. Before you install any skill, the same checks apply:
- Read the SKILL.md and any referenced scripts
- Check the GitHub repo's last commit date, star count, open issues
- Look for any network requests in the skill, and what does it call
- Look for any file writes outside the project directory
- Check what API keys or credentials the skill asks for
For a full process, see our skill audit checklist.
FAQ
Do I need to know how to code? Mostly no. Installing a skill is one command. Using it inside Claude Code is conversational. The exception is vexa's self-hosted setup, which assumes you're comfortable with Docker.
Can I run these skills without Claude Code? Some (elevenlabs-mcp, mcp-server-youtube-transcript) are MCP servers that work with any MCP client like Cursor, Cline, or custom apps. Others (deep-research, claude-seo) are Claude-specific skills that need the Claude Code runtime.
What about voice rights when generating intros? Your own voice, cloned with your own consent through ElevenLabs' verification: legally fine. Anyone else's voice: needs explicit written consent, and even then state laws vary. When in doubt, hire a voice actor for $50 instead.
How do these handle non-English podcasts? elevenlabs-mcp supports 30+ languages with quality varying by language. deep-research handles English best; other languages depend on how much source material exists. vexa transcription works in 50+ languages via Whisper. claude-seo's optimization patterns translate but you'd want a native speaker to review.
Are these going to break my existing workflow? No. Skills run inside Claude Code and don't touch your DAW, hosting platform, or distribution. They sit alongside what you already have.
What's missing from this list? Audio editing skills with real waveform manipulation. The Claude ecosystem hasn't caught up to Descript or Adobe Podcast on that front. For now, skills handle the workflow around production; tools handle the production itself.
Share this article
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.