TwelveLabs opens Jockey to everyone and turns video archives into usable content

TwelveLabs launches the public version of Jockey, its AI agent capable of searching, analyzing, and editing videos from ChatGPT or Claude.

Locating a line spoken in the middle of an interview, identifying all shots featuring a product, or assembling the best moments from multiple videos. With the general availability of Jockey, TwelveLabs brings these operations together into an agent specialized in audiovisual understanding. Accessible from ChatGPT and Claude, the system allows users to query video and photo libraries in natural language, review corresponding clips, and build exportable MP4 edits. Archives Searchable from ChatGPT and Claude

Jockey was already available in an experimental version, notably within Claude. Its public launch comes with integration into ChatGPT, via the MCP (Model Context Protocol).

Users can connect their TwelveLabs account, select a collection, and query their videos directly within a conversation.

It thus becomes possible to ask the assistant to find all clips in which a product appears, the moments when a speaker brings up a specific topic, or sequences showing a precise action.

Results include relevant clips and their timestamps, allowing users to locate the passages in the original files.

Jockey operates as a specialized layer that ChatGPT or Claude can call upon. Videos remain organized in the TwelveLabs library, while the assistant serves as the conversational interface.

The integration is documented in the TwelveLabs MCP connection guide. Searching for a Scene Rather Than a File Name

One of Jockey's main functions involves semantic search within audiovisual content.

Instead of relying on file names, folders, or manually added keywords, the system examines what appears and happens in the videos.

A search can target an object, a person, a place, an action, or a freely described situation.

A creator with hundreds of clips can, for instance, ask to find shots where a character crosses a street in the rain, even if no file contains those terms in its name.

Jockey relies on TwelveLabs technologies, notably Marengo, dedicated to multimodal representation and search, and Pegasus, designed for content and context analysis.

These models make it possible to match a text prompt with video passages and provide descriptions of the identified events.

Matches remain subject to the models' interpretation capabilities. An ambiguous scene or a hard-to-distinguish event may yield incomplete or incorrect results. Locating Spoken Words in an Interview

Searching is not limited to visual elements.

Jockey can also use spoken content to identify relevant passages in a recording.

An editorial team can ask to find all clips in which multiple people explain why they switched software, mention a difficulty, or share their thoughts on a product.

The agent leverages transcripts to identify corresponding speaking segments and allow them to be viewed within the original videos.

This feature can be used for editing interviews, preparing documentaries, or repurposing lectures and filmed podcasts.

Timestamps eliminate the need to scrub through entire recordings to find a quote.

However, accuracy depends on audio quality, transcription accuracy, and the match between the query and what was actually said. Breaking Down a Video into Scenes Directly in the Conversation

Jockey also allows users to request a scene-by-scene analysis.

Users can submit a video and ask for a breakdown of its progression. The agent then produces a breakdown that includes visual previews, descriptions, and links to the relevant clips.

These results can be viewed directly within ChatGPT or Claude, without needing to open each file separately.

For a product demo, for example, this makes it possible to identify the different steps presented, changes in environment, or moments when a feature appears on screen.

In video editing, this breakdown can facilitate selecting which clips to keep or move around.

It can also be used to examine the structure of existing content before producing a shorter version. Importing Videos from Multiple Sources

The public version simplifies adding content to TwelveLabs collections.

Users can submit a public YouTube link or a direct URL to a video file accessible online.

Multiple links can be provided in the same conversation, along with an instruction requesting their addition to a specific collection.

Importing from a computer is handled via an upload link provided by Jockey. This opens a page allowing files to be added to the selected library.

The announcement also mentions importing from Google Drive among the new capabilities offered.

Once content is processed and indexed, it becomes available for search and analysis.

This workflow involves a media preparation phase. Files are not necessarily searchable immediately after uploading, especially when dealing with large volumes. Building an Edit from Retreived Clips

Jockey can use search results to assemble multiple clips into a single video.

A user can request a selection of moments focused on a topic, a product, or a person, and then obtain an initial sequence built from the identified clips.

The tool now allows users to maintain precise in and out points, down to a fraction of a second.

TwelveLabs provides the example of a clip selected between 12.4 and 16.8 seconds. These boundaries are preserved during assembly, rather than being rounded to whole seconds.

This precision can notably prevent cutting off the end of a sentence or starting a clip after an action has already begun.

The result can then be opened in a timeline to continue editing. Reordering Shots Before Exporting

The new interface allows users to modify clip order and duration.

Users can shorten a passage, move a shot to use it as an opening, or scrub through the sequence to check its flow.

These operations remain separate from the initial search. The agent can suggest a first selection, while the final edit can be adjusted manually.

Once the video is prepared, Jockey allows it to be exported in MP4 format.

TwelveLabs also offers the ability to share certain results as artifacts. These elements retain links to the videos used, enabling another person to