Make videos in Claude Code
Connect Idea to Motion to Claude Code and you can ask for a video the same way you ask for anything else. Claude Code writes the script, plans each scene and writes the animation code; we synthesize the voiceover, time it to the words, render the animation and hand back an MP4. It speaks Streamable HTTP directly, so connecting is one config entry carrying an API key in an Authorization header.
- 1Mint a connection key below.
- 2Add the config entry to your client and reload it.
- 3Connect a voice key, then ask for a video in chat.
01Connect Claude Code
Generate a connection key if this client needs one, then add the server.
Pick your app
Chat apps
Editors & IDEs
Terminal
Generate a connection key
Add it to Claude Code
claude mcp add ideatomotion https://api.ideatomotion.com/mcp --transport http --header "Authorization: Bearer vai_YOUR_KEY"
Generate a key above to replace vai_YOUR_KEY.
Where the config lives
The exact names Claude Code reads, and what happens when one of them is wrong.
Claude Code reads its MCP servers from a JSON file — ~/.claude.json. It can live in either of two places, and the per-project one only applies inside that project. The entry goes under mcpServers, with the endpoint in url. Those names are not what every other client uses, and getting one wrong is not an error, it is silence: the file parses, Claude Code starts, and the server never appears.
- Config files
- ~/.claude.json (global) or .mcp.json (per project)
- Format
- JSON
- Root key
- mcpServers
- Endpoint field
- url
- Your key
- Written into the file as an Authorization header
- Fastest route
- The claude mcp add command, or the file below
- After saving
- Reload or restart Claude Code — MCP config is read when the client starts
02Connect your voice
Narration is synthesized on your own ElevenLabs, Cartesia or Sarvam AI key. Silent motion graphics need none.
Your voice key
Idea to Motion generates the voiceover on your own text-to-speech account. Add an ElevenLabs, Cartesia or Sarvam AI key — it's stored encrypted and never shown again. Rendering is blocked until a key is connected.
What Claude Code can actually do
Where the line between your assistant and the platform falls, and why it falls there.
Ask for a video in a Claude Code session the same way you ask for a refactor — the tool calls run inline and you approve the plan without leaving the terminal.
- It writes the content — a narrated script, a silent shot list, or a keep-and-drop cut plan for a video you upload — and hands it back for approval. Nothing is produced until you say yes.
- It directs every scene, then writes a self-contained Remotion component for each one. Submitted code is validated and run in a sandbox before it is accepted, so a mistake comes back as a tool error it can fix rather than a broken video.
- It renders stills of its own scenes on our servers and looks at them, which is why it can correct a card sitting across someone's face without you screenshotting anything.
- We do the parts a language model should not be doing: voice synthesis, word-level timing, captions, music, assembly and the final render.
- There is deliberately no export tool. Exporting spends money and picks a quality tier, so it stays a click you make on the platform — and a video you have already paid for re-exports for free.
A worked example
Type it at the prompt, in any directory — nothing about this needs a project.
“Make a 40-second widescreen explainer about what this CLI does, in the schematic look.”
What comes back
- 1It opens a picker: the looks play as short clips, the narrator voices have samples you can hear, and you choose. Clients that cannot draw a panel get the same list as text.
- 2It writes the script and shows it to you. Tell it what to change and it rewrites — the video does not start until you approve.
- 3It directs and writes each scene, checking stills as it goes. This is the long part, and it is mostly quiet: a scene submission legitimately holds its request open while the server renders.
- 4The finished video appears in your projects here. Watch it, ask for adjustments in the same conversation, then export it from the platform.
Troubleshooting Claude Code
The ways this goes wrong, and most of them go wrong silently.
A scene submission sits there for a minute or more
That call renders the scene on our servers to check it before accepting it, so it holds the request open for tens of seconds — longer behind a queue. It emits progress notifications every few seconds to keep your client's timer alive, so a quiet pause is the tool working, not a hang. Claude Code can also run one scene per subagent, which is the workflow the server hands it on connect.
Claude Code starts fine and the server never appears
Check the names in ~/.claude.json. The entry belongs under mcpServers, and the endpoint goes in url. These differ between clients — some read mcpServers, some servers, some context_servers — and a wrong one is not an error: the file parses, Claude Code starts, and the server is simply never loaded. Copy the snippet in step 1 rather than adapting one from another client's documentation.
It says it cannot connect, or every tool call comes back 401
A key that is wrong, truncated or revoked gets a 401, and most clients report that as a failed connection rather than as an auth problem. Keys are shown once and stored hashed, so there is no way to read an old one back — generate a fresh one in step 1, then re-run the claude mcp add command with the new key and reload Claude Code. Paste the whole vai_ string including the prefix; the old key can be revoked from the same panel.
You saved the config and nothing changed
This can live in two places: ~/.claude.json (global) and .mcp.json (per project). A per-project file only applies while you are working in that project, so an entry in the wrong one reads exactly like a config Claude Code is ignoring. Put it in the user-level file if you want it everywhere, and reload Claude Code either way.
The tools work but it cannot produce the voiceover
Narration and transcription run on your own ElevenLabs, Cartesia or Sarvam AI key rather than ours, so a narrated or recorded video needs one connected — step 2 on this page, or the set_voice_key tool from inside Claude Code. Silent motion graphics need no voice key at all, so "make it a silent motion graphic instead" is a valid way past this.