Aryankeluskar Poke Video MCP

What aryankeluskar-poke-video-mcp MCP server does

aryankeluskar-poke-video-mcp MCP server lets an MCP client search a personal Flashback video library using ordinary language. Instead of requiring filenames or timestamps, you can describe the content you want, such as a meeting about deadlines, cooking in a kitchen, or a red car driving.

The server returns formatted search results for matching video segments. Each result can include an AI-generated description, a relevance score from 0 to 1, a direct clip URL, and the URL expiration time. The clips are approximately 30 seconds long. Descriptions may cover both visible content and spoken audio because the underlying processing includes visual analysis and transcription.

How it works

Flashback processes uploaded videos into segments of about 30 seconds. It analyzes the visual material, transcribes audio, and stores searchable representations in a vector database. When a client submits a query, the service converts the text into embeddings and compares it with the indexed video segments. Results are ranked by semantic similarity rather than by an exact keyword match.

The aryankeluskar-poke-video-mcp MCP server interfaces with this processing system through its API. The documented backend uses FastAPI on Modal, Pinecone for vector search, Anthropic Claude for visual analysis, OpenAI Whisper for transcription, and Google Cloud Storage for video files. These backend details describe the service architecture; they are not separate MCP tools exposed to the client.

Setup and configuration

You need a Flashback account user ID and access to the Flashback video processing system. When adding the server to Poke or another MCP client, provide the user_id value associated with the Flashback account. The README shows a UUID-formatted example, but each deployment should use its own account identifier.

For local development, the documented commands are:

• uv run dev to start the server
• uv run playground to test it interactively

The repository also describes deployment through Smithery after pushing the project to GitHub. No environment variable names are specified in the provided material.

Tools and capabilities

The primary search operation is query_videos(query: str, max_results: int = 10). The query is a natural-language description of the desired content. max_results controls how many matches are returned and accepts values from 1 through 15. The listing may expose this operation under the shorter name query.

The server also documents get_setup_instructions(), which returns setup and usage guidance. In addition, it provides the api://video-processing resource containing information about the underlying video processing API.

Limitations and notes

Search only covers videos that have been uploaded to the relevant Flashback account. An empty result can therefore indicate that the collection does not contain matching or indexed material. Older videos may lack descriptions until they are processed again.

Clip URLs are presigned for security and expire after one hour. Ask for fresh search results when a previously returned URL no longer works. The supplied documentation does not specify client-specific compatibility beyond mentioning Poke and other MCP clients, nor does it document authentication fields other than user_id.

For developers evaluating the aryankeluskar-poke-video-mcp MCP server, the main fit is agent-driven retrieval from an existing Flashback collection. It is not documented as a general-purpose video editor, uploader, or transformation service.