dinoctor-mcp
# Dinoctor MCP server
Let Claude, Cursor or any [MCP](https://modelcontextprotocol.io) app direct animated videos in [Dinoctor](https://dinoctor.com), a free desktop video app for Mac and Windows.
You ask in plain words ("make a 30 second video about a dino who loses his sunglasses"), and the AI builds it in the open Dinoctor window while you watch: script, scenes, shots, places, characters, actions, camera moves, lights, on-screen text, voice, music, and the final MP4.
This repo is the MCP server only (the remote control). The video app itself is a separate free download.
## What you need
1. **Dinoctor**, free: [dinoctor.com](https://dinoctor.com) or the [downloads page](https://github.com/Dinoctor/dinoctor-downloads).
2. **Node.js 18 or newer**, if you use `npx` below. (Or skip Node: the app has this same server built in, see [Without Node](#without-node).)
## Set it up
### Claude Desktop
Open Settings, Developer, Edit Config, and add:
```json
{
"mcpServers": {
"dinoctor": {
"command": "npx",
"args": ["-y", "dinoctor-mcp"]
}
}
}
```
Restart Claude Desktop.
### Claude Code (plugin: server plus a directing skill)
The plugin adds the server and a skill that teaches Claude how to direct a good video: plan a hook, pick ids from the style's library, act, move the camera, check every shot with `look`, and always ask before spending credits.
```
/plugin install dinoctor --marketplace Dinoctor/dinoctor-mcp
```
On older Claude Code: `/plugin marketplace add Dinoctor/dinoctor-mcp`, then `/plugin install dinoctor@dinoctor`.
Server only:
```bash
claude mcp add dinoctor --scope user -- npx -y dinoctor-mcp
```
### Cursor
[](https://cursor.com/en/install-mcp?name=dinoctor&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsImRpbm9jdG9yLW1jcCJdfQ==)
Or add to `~/.cursor/mcp.json` (or your project's `.cursor/mcp.json`):
```json
{
"mcpServers": {
"dinoctor": {
"command": "npx",
"args": ["-y", "dinoctor-mcp"]
}
}
}
```
### Other MCP apps
Any app that runs stdio MCP servers works: the command is `npx -y dinoctor-mcp`.
### Without Node
The Dinoctor app is also the server: in Dinoctor open **Help, Connect to Claude or Cursor...** and click **Add to Claude Desktop**, or point your app at the app itself with the `--mcp` argument:
- Mac: `/Applications/Dinoctor.app/Contents/MacOS/Dinoctor --mcp`
- Windows: `%LOCALAPPDATA%\Programs\Dinoctor\Dinoctor.exe --mcp`
Use one way, not both, or your AI app will see every tool twice.
### The directing skill in other apps
The skill is a plain file: [skills/dinoctor-director/SKILL.md](skills/dinoctor-director/SKILL.md). In Claude Desktop or claude.ai, zip the `dinoctor-director` folder and upload it as a skill in Settings. In other apps, paste it into the project's rules or instructions.
## How it finds Dinoctor
The server talks to the Dinoctor app on your own computer only (127.0.0.1). It never sends anything to the internet by itself.
When the AI calls `open_dinoctor`, the server starts the app if it isn't open:
- Mac: `/Applications/Dinoctor.app` or `~/Applications/Dinoctor.app`
- Windows: `%LOCALAPPDATA%\Programs\Dinoctor\Dinoctor.exe` or `%ProgramFiles%\Dinoctor\Dinoctor.exe`
- Somewhere else (or the Microsoft Store version): set `DINOCTOR_APP` to the full path of the app.
| Setting | What it does |
| --- | --- |
| `DINOCTOR_APP` | Full path to the Dinoctor app, if it isn't in the usual place. |
| `DINOCTOR_PORT` | The port of a Dinoctor that is already open, if you changed it. Normally found by itself. |
| `DINOCTOR_HOME` | Dinoctor's data folder, if you moved it. |
Without the app, the server still starts and lists its tools; calls then answer that Dinoctor isn't open.
## What the AI can do
69 tools, in groups:
- **Start and read**: `open_dinoctor`, `status`, `get_project`, `get_shot`, `go_to`, `how_to`
- **Build the video**: `new_project`, `set_style`, `set_format`, `add_scene`, `add_shot`, `remove`, `move`, `duplicate`, `cut_shot`, `rename_scene`, `edit_project`
- **Direct a shot**: `cast`, `act`, `words`, `camera`, `light`, `place`, `keys`, `transition`
- **See it**: `list_styles`, `library`, `look` (a picture of any moment, so the AI checks its own work)
- **Script, voice and sound**: `write_script`, `get_script`, `assign_line`, `voices`, `speak`, `own_voice`, `music`, `sound_effects`
- **Pictures**: `generate_picture`, `generate_character`, `generate_place`, `picture_status`, `generated_pictures`, `put_picture`, `remove_background`, `relight`, `cutout`, `upload`, `generator_account`
- **Styles and starting points**: `my_styles`, `make_style`, `style_status`, `continue_style`, `fix_style`, `cancel_style`, `rate_style`, `delete_style`, `use_style`, `samples`, `fill_spot`, `templates`, `describe_drawing`, `draw_character`, `draw_object`, `draw_place`
- **Files and export**: `render`, `render_status`, `cancel_render`, `list_projects`, `open_project`, `save_project`
## Safe by design
- **Nothing paid happens without you.** Tools that would spend credits (AI pictures, ElevenLabs voices, Claude usage for building styles) first answer with the price. They only go ahead when called again with `confirm: true`, and the AI is told to ask you first.
- **Everything can be undone.** Each change the AI makes is one undo step in Dinoctor, with a note saying what it did.
- **Local only.** The server only talks to the Dinoctor app on your computer.
## Try it
Open your AI app and say:
> Open Dinoctor and make a 20 second pixel-style video: a dino walks into a desert, finds a cactus, and jumps for joy. Look at each shot and fix anything that looks off.
## Develop
```bash
npm install
npm test # starts the server and lists its tools (works without the app)
node test/list-tools.mjs status get_project # call tools against an open Dinoctor
```
## License
MIT. This license covers this MCP server only. The Dinoctor app is free to use but is not open source.
TDQS
Scored across 69 tools
Most tools map to distinct resources or actions, and many descriptions explicitly redirect to the right alternative. However, several overlapping families create real ambiguity: use_style vs. set_style, generate_* vs. draw_* vs. upload vs. put_picture, speak vs. own_voice vs. assign_line/write_script, and multiple ways to open content (open_project, templates open, samples open, new_project). The detailed cross-references help, but the 69-tool surface is still hard to navigate.
All names use snake_case, so casing is consistent and readable. But the convention mixes noun-only tools (voices, music, cast, act, words, camera, light, place, keys, transition, library, samples, templates, status, look) with verb-based tools (get_project, add_shot, remove, generate_picture, cancel_render), so there is no predictable verb_noun pattern.
The domain is genuinely broad — project editing, style building, AI picture/voice generation, rendering, audio, scripts, samples and templates — so some breadth is justified. Still, 69 tools is far beyond a typical well-scoped set and likely overwhelms tool selection; it would benefit from grouping or splitting into narrower servers.
The surface covers nearly the full lifecycle: project open/save/list, scene/shot CRUD, casting, acting, text, camera, lighting, places, keyframes, transitions, script and voice handling, sound/music, rendering, style management, generation, uploads, samples and templates. Minor gaps remain, such as no explicit undo/redo tool, no project deletion, and limited deletion/management for media or generated pictures.