Overview
Agents use the Jumper MCP server to interact with Jumper. This means that the agent only sees the data that is returned by the MCP server. Your footage never leaves your computer. Jumper does all analysis locally on your machine. What the agent receives is metadata and derived data — not your media files.What Jumper’s MCP returns
When querying the JumperMCP server, depending on the task, the agent can receive:- Metadata — File paths, timecodes, clip durations, similarity scores
- Downsampled thumbnails — Low-res previews of search results so the agent can assess which matches fit your query (e.g. which “man in a red shirt” shot is best)
- Transcription excerpts — When the task involves speech search, the agent receives the transcribed text for the relevant segments
The privacy boundary
Data flows like this: Jumper controls what goes out. The agent only sees what the MCP endpoints return.Local vs remote models
Local model — If the agent runs a local model (e.g. Codex with a local LLM), the MCP responses never leave your machine. Remote model — If the agent uses Claude API or OpenAI, the tool output becomes part of the prompt sent to the provider. That means metadata, thumbnails, and transcript excerpts you return are processed on their servers. Same as pasting that data into ChatGPT.Model provider policies
Each provider has its own data handling:- Anthropic (Claude) — Commercial customers: no training on your data unless opted in. Enterprise can enable Zero Data Retention (ZDR). Standard retention: 30 days. Note: data processed by third-party tools and MCP servers is not covered by ZDR; review each integration.
- OpenAI — Check their API data usage policies for retention and training.