videoctx
Read videos through a local text index, with frames fetched only when needed
- Pricing
- Free
- Platforms
- CLI, macOS, Linux
- Runs
- Locally
- Version
- 0.1.5
- Listing updated
- today
Similar MCP servers
About videoctx
videoctx is a free MCP server for files and documents. It runs locally on your machine and is started by your AI client.
videoctx builds an open, portable index of a video, stored as a .vctx folder of JSON and JPEG, using only local tools: ffmpeg keyframes, PySceneDetect shot boundaries, faster-whisper transcripts and optional one-line captions. A question is answered from the text layer first, and individual frames are fetched only when it needs pixels.
The README argues this avoids re-tokenizing an hour of video on every question, quoting roughly 250 to 350 thousand tokens, and reports about two cents for a follow-up on an hour-long video. Its Video-MME results are described with caveats: on short videos the latest version scored 84.7 percent, and medium-length videos are called its weakest case.
The MCP server makes no model calls and needs no key; the calling agent reads the text and frames itself. A command-line ask feature uses the Anthropic API. It is aimed at developers comfortable with Python and ffmpeg.
Install
Run in a terminal
claude mcp add videoctx -- uvx videoctxAdd to claude_desktop_config.json
{
"mcpServers": {
"videoctx": {
"command": "uvx",
"args": [
"videoctx"
]
}
}
}Add to ~/.cursor/mcp.json
{
"mcpServers": {
"videoctx": {
"command": "uvx",
"args": [
"videoctx"
]
}
}
}Add to .vscode/mcp.json
{
"servers": {
"videoctx": {
"type": "stdio",
"command": "uvx",
"args": [
"videoctx"
]
}
}
}Add to ~/.codeium/windsurf/mcp_config.json
{
"mcpServers": {
"videoctx": {
"command": "uvx",
"args": [
"videoctx"
]
}
}
}Add to your client's MCP settings (most use this shape)
{
"mcpServers": {
"videoctx": {
"command": "uvx",
"args": [
"videoctx"
]
}
}
}Using several servers? Build one config file with the MCP config generator.
Tools this server provides (6)
Taken from the project documentation; the running server may differ.
- video_overview
- Cheap text-layer overview of a video
- survey_video
- One thumbnail per shot
- get_frames
- Fetches frames between two times at a chosen detail level
- transcript_search
- Searches the transcript
- find_moments
- Semantic visual search, needs an embed step
- index_video
- Indexes a video
What it does
- Local indexing
- Indexes video offline with free tools.
- Open format
- Plain JSON and JPEG .vctx folders.
- Text-first queries
- Answers from transcript and captions before fetching frames.
- On-demand frames
- Pulls specific frames for visual questions.
- Semantic moment search
- Finds visual moments after a one-time embed step.
- Published benchmarks
- Shows accuracy and cost with caveats.
Good for
- Summarise a lecture with timestamps
- Find action items in a recorded meeting
- Locate an error dialog in a screen recording
- Index a talk archive once and query it repeatedly
Pricing
Free
Indexing is local; the optional ask command uses your Anthropic API key
videoctx FAQ
Does indexing cost anything?
The README says indexing is local and free.
Does the MCP server need a key?
No, it makes no model calls.
What do I need installed?
Python and ffmpeg.
What is the cost of follow-up questions?
About $0.02 for an hour-long video, per the README.
Is videoctx free?
Yes. videoctx is free to use.
How do I add videoctx to Claude Code?
Run this command in your terminal: claude mcp add videoctx -- uvx videoctx
What are the best alternatives to videoctx?
The closest alternatives to videoctx are headroom, omniget, next-ai-draw-io, mcp-for-blender and Claude Code Ultimate Guide.
Where this comes from
- Description: written from the project's README.
- Install steps: from the official MCP Registry entry.



