← All experiments

EchoDeck

Since Claude can build PowerPoints, I wanted to see if I could give it my voice to do a full on presentation.

See what I did, then go ahead and add your voice to your Claude account today.

How it works

1
Ask Claude to build a deck
"Can you make me a powerpoint on some cool new topic, make it 2 slides and include speaker notes"
Asking Claude to create a PowerPoint presentation
2
Claude builds the entire presentation
It picks a topic (Neuromorphic Computing), writes the slides, adds speaker notes, and hands you a downloadable .pptx file.
Claude presenting the finished PowerPoint with download
3
Ask Claude to generate a video — in your voice
Claude finds the EchoDeck MCP tools, discovers your cloned voice, and starts generating.
Claude using EchoDeck to generate a narrated video
4
Video is ready
1080p, 1 minute 7 seconds, with crossfade transitions between slides and Kunal's voice reading the speaker notes.
Finished video ready for download
5
The output
This video was generated entirely by Claude + EchoDeck. No recording. No editing. No microphone.

Try it yourself

Step 1: Get an ElevenLabs API key

Sign up at ElevenLabs and grab a free API key. This is what powers the voice synthesis — no credit card required to get started.

Step 2: Deploy EchoDeck

Deploy on Railway to spin up your own server. Railway will ask you to configure environment variables — all that is needed is your ELEVENLABS_API_KEY.

Step 3: Connect it to Claude

Once deployed, grab your server URL from Railway. Then:

  1. Open Claude Desktop → Settings → Integrations
  2. Add a new Custom MCP Connector
  3. Paste in your EchoDeck MCP URL

That's it. Claude now knows how to make videos.

Step 4: Build a presentation from scratch

Open a new Claude conversation and type:

"Build me a PowerPoint on [any topic you want] with speaker notes."

Claude will generate a full .pptx — slides, content, speaker notes — and hand you the file. No template. No outline. Just a finished deck.

Step 5: Turn it into a narrated video

In that same conversation, say:

"Turn this into a narrated video using the Bella voice."

Claude uploads the deck to your EchoDeck server, generates speech from the speaker notes, stitches everything into an MP4 with transitions, and gives you a download link. The whole thing takes about a minute.

You just went from an idea to a fully narrated video presentation without opening PowerPoint, a video editor, or a microphone.

A note on security: This MCP server has no authentication. Use it as an experiment only.

Technical challenges

Tool call limits

Claude polls the server to check if your video is finished rendering. On longer videos, it hits a tool call limit before the video is done. This is a Claude platform limitation — it happens with Claude's own built-in tools too — and will likely go away as Anthropic raises the ceiling. For now, longer videos will prompt you to manually tell Claude to "continue" until rendering completes.

Getting large files to the server

The harder problem was file transfer. MCP tools communicate through Claude's context window, which means you can't just pass a 10MB PowerPoint file as a tool argument — it would blow through the token limit.

The solution was a two-step handoff. Instead of sending the file through the MCP protocol, the get_upload_slot tool returns an upload URL and a pre-built curl command. Claude then uses its built-in command execution to run that curl command directly — uploading the file out-of-band, outside the context window entirely. The server responds with a file_id, and Claude passes that small string to generate_video. Large file in, tiny reference out.

View Source →