← Back to Podcli
MCP server · 19 tools · Open source

Clip podcasts from Claude Code

Podcli ships as a Model Context Protocol server with 19 tools that cover transcription, AI clip scoring, face-tracked 9:16 cropping, burned-in captions, and export. Plug it into Claude Code, Claude Desktop, or Codex. Say “clip this episode” and the agent runs the whole pipeline.

View on GitHubInstall in 30 seconds

You talk to the agent. The agent calls Podcli.

MCP is the spec that lets an AI agent discover and call tools on your machine. Podcli runs as one of those servers.

You
Natural language
“Transcribe episode-072 and queue 8 clips with karaoke captions.”
Agent
Claude / Codex / Cursor
Reads Podcli's 19 tools, picks the ones needed for the request, and calls them in order.
Podcli MCP
Local process, 19 tools
Runs transcribe → score → crop → caption → export. The video file never leaves your machine.

The 19 tools, grouped by stage

Every step of the podcast-to-shorts pipeline is exposed as a tool. The agent picks the right ones automatically.

Ingest

transcribe_podcast
Run Whisper plus speaker diarization on a local file. Cached by file hash.
transcribe_start
Async transcribe with progress events.
set_video
Register a video without transcribing it.
import_transcript
Import an existing transcript (Descript, Otter, plain text).
parse_transcript
Parse plain text into word-level timestamps.

Score

suggest_clips
Hand transcript plus knowledge base to Claude or Codex, get scored candidates.
analyze_energy
Find high-energy moments by audio analysis.
modify_clip
Adjust start, end, title, or caption style on a candidate.
toggle_clip
Mark candidates as selected or skipped.
knowledge_base
Read or update the show's voice, banned words, title formulas.
clip_history
Track shipped clips and check for duplicates against the episode DB.

Render

create_clip
Render one clip with face tracking, 9:16 crop, captions burned in.
batch_create_clips
Render every selected candidate in one pass.
manage_presets
Save and load render configs (caption style, crop, logo).
update_settings
Change caption style, crop, logo for the active session.
manage_assets
Register logos and overlays by name.
list_outputs
List the rendered files for the session.

Ops

job_status
Poll any async job (transcribe or render).
get_ui_state
Read the current session state and transcript.

Add Podcli to Claude Code

The MCP config is one block. Same shape for Claude Code, Claude Desktop, and Codex.

// claude_desktop_config.json (same shape for Codex)
{
  "mcpServers": {
    "podcli": {
      "command": "node",
      "args": ["dist/index.js"],
      "cwd": "/path/to/podcli"
    }
  }
}

What you actually say to the agent

Natural language. The agent reads the tool list and figures out the call sequence.

›Transcribe episode-072.mp4 and suggest the top eight clips.
›Render the selected clips in 1080×1920 with the karaoke caption style.
›Check clip 4 against the episode database. Have we shipped anything close to this before?
›Update the knowledge base to add 'monetization' as a banned topic for this season.
›Batch-process every .mp4 in ./episodes/season-3/ at top=10.

Why not just use a generic FFmpeg MCP

Same prompt, two tool surfaces. The agent does less work when the tools speak podcast, not pixels.

Generic FFmpeg MCP
Low-level primitives
trimcropconcatscaleencode
Agent has to invent: where moments are, how to crop a face, what burned-in captions look like, dedupe against past clips.
Podcli MCP
Podcast-clipping primitives
suggest_clipscreate_clipknowledge_baseclip_historyanalyze_energy
Agent reasons about clips, not FFmpeg flags. “Find the 8 best” is one tool call.

Frequently asked

The questions people search before they wire Podcli into their MCP client.

What is the Podcli MCP server?+

A Model Context Protocol server that exposes 19 tools covering the podcast-to-clip pipeline: transcription, AI clip scoring against a knowledge base, face-tracked 9:16 cropping, burned-in captions via Remotion, hardware-accelerated export.

How do I add Podcli to Claude Code?+

Install Podcli (git clone, ./setup.sh), then point your MCP client config at the built server. The same JSON works for Claude Desktop and Codex.

Does the MCP server upload my podcast files anywhere?+

No. It runs locally as a process the agent talks to. Video, transcript, and renders all stay on your machine. The only optional network calls are to the Claude or Codex API for AI clip scoring.

Is this different from a CLI?+

Yes. The CLI is a fixed command. The MCP server exposes discrete tools the agent composes on demand. Both ship in the same install and you can use either.

What models work with Podcli's MCP server?+

Any MCP-compatible client. Tested with Claude Code, Claude Desktop, and Codex. Any agent that speaks MCP can read the tool list and call it.

Wire Podcli into Claude Code

Clone the repo, run the setup script, then point your MCP client at the built server. Takes about as long as making coffee.

$ git clone https://github.com/nmbrthirteen/podcli.git
$ cd podcli
$ ./setup.sh
View on GitHubOpen-source OpusClip alternative