Podcli - AI Podcast Clipper with Face Tracking and MCP

TL;DR

TL;DR: Podcli is an open-source AI podcast clipper that runs entirely locally — transcribes episodes, picks the best moments using AI or audio energy, crops to the speaker with face tracking, and burns in captions. Ships with 26 MCP tools so a coding agent can drive the full clip workflow.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: podcli.com ← verified
• Source repository: github.com/nmbrthirteen/podcli ← verified
• License: AGPL-3.0 ← verified from LICENSE file
• HN launch thread: news.ycombinator.com/item?id=47209491 ← verified
• Stars (as of 2026-07-17): 25 (main branch)
• Source last checked: 2026-07-17 (commit main)

What Is Podcli?

Podcli is an open-source AI podcast clipper built for creators who want full control over their workflow. It runs entirely on your machine — transcription and rendering are local by default. The only outbound calls are the optional AI scoring engine (Claude/Codex) and YouTube publishing.

The tagline from the README:

“Open-source AI podcast clipper. Turn a long episode into short clips with face tracking and burned-in captions. Drive it from the CLI, a web studio, or your coding agent.”

Setup Workflow

Step 1: Install

No prerequisites required. The installer fetches a self-contained binary and provisions Python, Node, FFmpeg, whisper.cpp, and the models it needs on first run.

macOS and Linux:

curl -fsSL https://podcli.com/install.sh | sh

Windows (PowerShell):

irm https://podcli.com/install.ps1 | iex

Runs on macOS (Apple Silicon), Linux (x64 and arm64), and Windows (x64).

Step 2: Process Your First Episode

podcli process episode.mp4

This transcribes the episode, picks the best moments, crops to the speaker, and renders clips with captions. Clips land in podcli-clips/ in the directory you ran it from.

To open the web studio:

podcli

The studio runs at localhost:3847 and includes a library, episode workspace, per-clip detail, highlights, thumbnails, content, analytics, assets, knowledge base, config, integrations, and MCP setup.

Step 3: Connect an AI Agent

Podcli is an MCP server with 26 tools. To register it with Claude Code:

podcli mcp install

Once registered, an agent can transcribe, suggest clips, render them, and publish to YouTube — all through conversation.

Deeper Analysis

What You Get

Clip formats: 9:16, 16:9, or 1:1, with captions sized per canvas. Face tracking follows the speaker, including split-screen layouts. Multi-segment cuts drop filler, pauses, and tangents automatically.

Caption styles: Four options — branded, hormozi, karaoke, and subtle.

Finding the moments: Whisper transcription with speaker diarization is the default. You can also bring your own transcript as .txt, .srt, or .vtt. AssemblyAI is available as an alternative engine. yt-dlp pulls episodes directly from a URL.

AI scoring matches potential clips against your knowledge base and episode history — it learns what you have already published and stops resuggesting those moments.

Hardware acceleration: VideoToolbox (macOS), NVENC (NVIDIA), and VAAPI (Linux) for real-time encoding, with a CPU fallback.

Publishing: Direct upload to YouTube with performance analytics to track which clips performed best. Davinci Resolve export as FCPXML for finishing by hand.

Comparison to Cloud Clippers

The README draws a direct comparison:

• Local by default — transcription and rendering happen on your machine; only optional AI scoring and YouTube publishing send data out
• Free and open source — AGPL-3.0, unlimited exports, full quality, no watermark
• Agent-native — 26 MCP tools let Claude Code or Codex drive the full flow from transcription through publishing
• Knowledge base — keeps titles, captions, and descriptions in your show’s voice, and prevents resuggesting published moments
• Davinci Resolve handoff — export any clip as FCPXML for manual finishing

Content Workflow with PodStack

PodStack ships with Podcli as a set of Claude Code slash commands. The workflow takes a transcript to a publish-ready package: scored moments, titles, descriptions, thumbnail briefs, a brand review, and a publish checklist.

Practical Evaluation Checklist

• Local transcription and rendering (no cloud required for core workflow)
• Face tracking that follows the speaker in frame
• Speaker diarization in Whisper output
• Four caption styles (branded, hormozi, karaoke, subtle)
• Multiple output formats: 9:16, 16:9, 1:1
• Hardware encoding via VideoToolbox, NVENC, or VAAPI
• MCP server with 26 tools for agent-driven workflows
• YouTube publishing with analytics
• Davinci Resolve FCPXML export
• Knowledge base that prevents re-suggesting published clips
• yt-dlp integration for pulling episodes from URLs
• Self-contained installer (fetches all dependencies)

Security Notes

• Local-first: All transcription and rendering run locally. No episode data leaves your machine unless you explicitly use cloud AI scoring or publish to YouTube.
• MCP access: The MCP server exposes tools that can transcribe, render, and publish. Restrict access to trusted agents only.
• YouTube OAuth: Publishing requires a YouTube account connection — treat the credentials as sensitive.

FAQ

Q: Does it require an internet connection? A: No — the core workflow (transcription, clip selection, rendering, caption burning) runs entirely offline. Only optional AI scoring with Claude/Codex and YouTube publishing require internet.

Q: What models does it use for transcription? A: whisper.cpp for local transcription by default, with speaker diarization. AssemblyAI is available as a cloud alternative.

Q: How does face tracking work without a dedicated AI model? A: The README does not specify the exact face tracking implementation. The workflow crops to whoever is speaking — likely using speaker diarization timestamps to drive the crop region.

Q: Can I use my own transcript instead of auto-transcription? A: Yes. Bring your own transcript as .txt, .srt, or .vtt.

Q: Is there a limit on clip exports? A: No. Exports are unlimited, full quality, and watermark-free per the AGPL-3.0 license.

Conclusion

Podcli fills a specific niche: the solo podcaster or small team that wants the power of AI clip selection and face tracking without surrendering their content to a cloud service. The MCP server integration is what sets it apart for developer-audiences — if you already use Claude Code or Codex, the 26-tool MCP interface lets you automate the entire pipeline from transcript to published clip.

The self-contained installer is the right call for this audience — one command gets you everything, including FFmpeg and whisper.cpp, without manual dependency management.

Source: podcli.com | GitHub | AGPL-3.0 | 25 stars