mediasense

Give Claude ears and eyes. Claude can't listen to audio or watch videos on its own; with mediasense it can. Point at a voice message, podcast, recording or meeting, or paste a TikTok, YouTube or Instagram link, in Claude Code or Claude Desktop and ask in plain language. Claude can then transcribe it, summarize it, explain what happens, find the moment something is said, or edit it.

you: what does this voice message say? ~/Downloads/audio-from-maria.opus
you: summarize this https://www.tiktok.com/@someone/video/7350…
you: what does he show at 0:30? and what tool is he talking about?
you: clean up ~/Desktop/podcast.m4a — remove the "um"s and long pauses
you: make short captions for ~/Videos/reel.mp4
you: put intro.mp3 under the first 5 seconds of talk.mp4, ducked

_{Real output for a 17-second TikTok by @elim.doc, shown with the creator's permission. Regenerate the animation with python docs/make_demo.py.}

mediasense runs on your computer: it downloads the link, transcribes speech in ~99 languages with a timestamp for every word, picks key frames at scene changes, and hands Claude a transcript, a shot list and contact-sheet images. Nothing is uploaded anywhere.

Português 🇧🇷

Setup

Windows – paste in PowerShell:

irm https://raw.githubusercontent.com/joaoestella/mediasense/main/install.ps1 | iex

macOS / Linux – paste in a terminal:

curl -LsSf https://raw.githubusercontent.com/joaoestella/mediasense/main/install.sh | sh

Then quit Claude Desktop completely (tray/menu-bar icon → Quit), open it again and ask "summarize this video: ". The first transcription downloads the speech model (≈ 500 MB).

irm https://raw.githubusercontent.com/joaoestella/mediasense/main/install.ps1 -OutFile install.ps1
notepad install.ps1
powershell -ExecutionPolicy Bypass -File .\install.ps1

winget install ffmpeg                    # macOS: brew install ffmpeg · Linux: sudo apt install ffmpeg
python -m pip install --user pipx
python -m pipx ensurepath                # then close and reopen the terminal
python -m pipx install "mediasense[all]"
mediasense setup -n                      # preview what setup will change
mediasense setup
mediasense doctor

/plugin install mediasense --marketplace joaoestella/mediasense

{ "mcpServers": { "mediasense": { "command": "uvx", "args": ["mediasense", "mcp"] } } }

What mediasense setup changes

• Claude Desktop: adds one entry, mcpServers.mediasense, to claude_desktop_config.json (%APPDATA%\Claude\ – or %LOCALAPPDATA%\Claude-Data\claude_…\ for the Microsoft Store app – ~/Library/Application Support/Claude/ on macOS, ~/.config/Claude/ on Linux; Claude Desktop → Settings → Developer → Edit Config opens the one your app uses). Everything else in the file is kept, and the previous file is saved next to it as claude_desktop_config.json.<date>.bak. If the file isn't valid JSON, setup stops without touching it.
• Claude Code (if installed): runs claude mcp add --scope user mediasense -- <path>\mediasense mcp.
• Nothing else. mediasense setup -n shows the exact changes without making them; --desktop or --code limits it to one app. To undo: remove the mediasense entry from the config file and run claude mcp remove mediasense --scope user.

Updating on Windows and pipx says a file "is being used by another process"? Claude Desktop is still running the old server: Get-Process mediasense | Stop-Process -Force, then install again (the installer script does this for you).

Something not working? mediasense doctor checks FFmpeg, the speech model, the GPU, link downloads and the Claude connection.

Teach Claude when to use it (recommended)

mediasense comes with a skill: a short guide that makes Claude reach for mediasense whenever you paste a video link, instead of trying to open the page (TikTok & co. block that).

• Claude Code: installed automatically by mediasense setup.
• Claude Desktop / claude.ai: download mediasense-skill.zip (the installer also saves it to Downloads/mediasense/), then in Claude open Settings → Skills and upload it. One time only; it syncs to all your devices.

Note: Claude Desktop sends files you attach in the chat to the model, not to your computer, so give mediasense a link or a file path instead (in Claude Code you can drag the file into the terminal).

Slash commands

The server adds ready-made prompts. In Claude Code they appear as /mcp__mediasense__summarize and so on.

Tools Claude gets

Long work never hits Claude's tool time limit: if a video takes more than ~30 s (a long video on a CPU, or the very first run while the speech model downloads), the tool answers "still working" with a progress percentage and Claude checks back with check_job until it's done.

All tools accept a file path or a link. Links are downloaded once to ~/Downloads/mediasense (change with MEDIASENSE_HOME) and reused.

Command line

Everything also works without Claude:

mediasense understand "https://youtu.be/…"           # transcript + shots + metadata, as text
mediasense transcribe talk.mp4 -l pt --vocab "Cloudflare, Claude" --also srt
mediasense search talk.transcript.json "next quarter"
mediasense edit talk.mp4 --fillers --silences 0.7 -n   # -n = only show the plan
mediasense edit talk.mp4 --fillers --silences 0.7 --srt --edl --fcpxml
mediasense edit talk.mp4 --cut-between "let me check" "okay so" --cut-range 12:03-12:40
mediasense export talk.transcript.json captions.srt --words-per-cue 2
mediasense place video.mp4 voiceover.wav --at 0:12.5 --mode duck
mediasense storyboard reel.mp4
mediasense fetch "https://www.tiktok.com/@…/video/…"

Edits never touch the original: results are written as <name>.edited.* / <name>.placed.*. EDL opens in Premiere, Resolve and Avid; FCPXML in Final Cut Pro and Resolve.

Good to know

• Names and jargon. Whisper may write "Cloud Fair" for Cloudflare. Pass --vocab (Claude does this automatically when it can infer the names from the title or the screen).

• Fillers. Whisper tends to leave hesitations out; transcribe with --verbatim when you want to cut them. Real words are protected per language ("um" is an article in Portuguese, so it is never cut there). Ambiguous words like "like", "tipo" or "né" are only removed if you list them.

• Speed. Everything works on a plain CPU: short clips take seconds, a few minutes of audio takes about as long as the audio with the default small model (base is faster, medium more accurate). No GPU? --backend openai sends the audio to an OpenAI-compatible transcription API instead (fast on any machine, but the audio leaves your computer).

• NVIDIA GPU (optional, much faster). mediasense uses it automatically once the CUDA libraries are present; no CUDA toolkit needed:

  python -m pipx inject mediasense nvidia-cublas-cu12 "nvidia-cudnn-cu12>=9,<10" nvidia-cuda-runtime-cu12
  

  (≈ 1.5 GB; needs a driver that supports CUDA 12 – mediasense doctor checks both). Restart Claude Desktop afterwards.

• Word timing is approximate (0.1–0.2 s). For tighter cuts and speaker labels use --backend whisperx (pip install "mediasense[whisperx]").

• Downloading from links follows each site's terms. Use it for content you're allowed to download.

Privacy Policy

mediasense runs entirely on your computer: no accounts, no analytics, no telemetry, and the author receives no data. It connects to the internet only to download the links you ask about, the speech model (first run), FFmpeg if it's missing, and – only if you choose the openai backend – to send audio to the transcription API you configure. Results stay on your disk until you delete them. When used from Claude, tool results (transcripts, key frames, metadata) are returned to the Claude app and handled under Anthropic's privacy policy. Full policy: PRIVACY.md.

Development

git clone https://github.com/joaoestella/mediasense && cd mediasense
pip install -e ".[dev,mcp]"
pytest

The tests build synthetic audio/video with FFmpeg and use fake model and downloader libraries, so they run in seconds with nothing to download. Ideas welcome: more filler words for your language, a one-click Claude Desktop extension (.mcpb), and a GUI.

Releasing

1. Bump version in pyproject.toml, src/mediasense/__init__.py, server.json (both places) and .claude-plugin/plugin.json, and the pinned mediasense==X.Y.Z in .mcp.json, commit. The workflow refuses a tag that doesn't match them.
2. git tag vX.Y.Z && git push origin vX.Y.Z. The Release workflow publishes to PyPI, updates the entry in the MCP Registry and creates a GitHub release with the wheel, mediasense-skill.zip and mediasense.mcpb.

One-time PyPI setup (trusted publishing, no token): on pypi.org → Your account → Publishing → Add a new pending publisher: project mediasense, owner joaoestella, repository mediasense, workflow release.yml, environment left empty.

License

MIT