Montage

Create a moment (clip) from [start, end] second-pairs of an ingested video. Saves the moment and starts clip preprocessing (downscale + speaker detection) in the background. Pick `segments` from the transcript / video-corpus timestamps, listed in ascending start-time order (the order they appear in the transcript); out-of-order pairs are sorted before the moment is saved. `kind` is one of content, hook_montage, ad, sponsor_read. `file_id` is only needed when the project has multiple files. Set `is_sizzle_reel` to true when the moment is a sizzle reel — a promo cut stitched from several non-contiguous highlights — rather than one continuous clip. `platform` is the intended destination: tiktok, youtube, instagram, linkedin or general (omit if unknown). Both are recorded on the moment's metadata and do not change `kind`. Speaker detection uses the recording's stored transcript, including when `utterances` is omitted. The legacy argument is accepted but ignored; it cannot override the recording's text, speaker IDs, or timestamps. Returns JSON with `moment.id`, `preprocessing.workflow_id` and `editor_url`. When `preprocessing.status` is "processing", poll `get_workflow_status(preprocessing.workflow_id)` until `completed` (typically a few minutes), then share `editor_url` with the user. A "skipped" status (see `preprocessing.reason`) means preprocessing could not run and there is nothing to poll; the moment still exists.

Create a new playbook from a natural-language description. Generates label, ranking, and hook rubrics via LLM based on your description, then saves the playbook as version 1 (auto-approved). `base_format_type` must be one of: default, podcast, talk. Returns the full playbook including generated rubrics.

Explain what Montage can do with recordings, including beyond clips. Use when a user asks what Montage can do or needs help choosing a use case. Returns examples, supporting tools and where results appear. Written synthesis is performed by the connected assistant; this tool does not read recordings, generate content or report account readiness.

Get moments for a project or file. At least one of `project_id` or `file_id` is required. Optionally filter by `kind`: content, hook_montage, ad, sponsor_read. Results are ordered by start time then creation time. Paginate with `limit` and `cursor` (pass the returned `next_cursor` as `cursor`). Moments with a downscaled clip include a presigned `video_url`, rendered as a playable gallery via the linked ui:// resource.

Get the current approved playbook by name. Returns the playbook's rubrics (label, ranking, hook) and metadata. Raises an error if the playbook is not found or does not belong to the user.

Get details for a single project by its ID. Returns the project's details if it exists and belongs to the authenticated user. Raises an error if the project is not found or does not belong to the user.

Get transcript segments for a project. Pass `file_id` or the exact `file_name` to select a specific file's transcription when the project has more than one file. Pass `provider` to select a specific transcription provider ('assemblyai' or 'deepgram'); defaults to assemblyai with automatic fallback to deepgram. Use `timestamp_format='hh:mm:ss:ms'` for human-readable timestamps instead of raw seconds. Paginate with `limit` and `cursor` (pass the returned `next_cursor` as `cursor` in the next call).

Get the progress of an `upload_video` pipeline. Returns JSON with `status` (running/completed/failed/...) plus, while running, `current_step` (creating_records/validating_url/downloading/ validating_video/analyzing/finalizing), `project_id`/`file_id` once created, and per-stage `child_workflow_status`. When completed, includes the final `result`; on failure, an `error`.

Get the vision-model video analysis for a project or file. Produced by the video-understanding pass of `upload_video`. Each result carries the `model` used, `latency_seconds`, `finish_reason`, the detected `category_name`, `max_people_count`, and the full `analysis` JSON — scene breakdown with timings and descriptions, clip candidates with hook/payoff reasoning, and reframe instructions. Read this alongside the transcript when choosing moment boundaries: the transcript gives the words, the analysis gives the visual story. At least one of `project_id` or `file_id` is required. Paginate with `limit` and `cursor` (pass the returned `next_cursor` as `cursor`).

Get the long-footage video corpus index for a project or file. At least one of `project_id` or `file_id` is required. `view="rollup"` (recommended first read) returns each result's bounded `rollup` summary — coverage fraction, subject roster with speaking/tracked seconds, hard-veto ranges, alignment-state totals, screen-text inventory — small enough to reason over many videos at once. `view="full"` (default) additionally returns the complete `corpus` index JSON (audio/video timelines, canonical `subjects`, constraints, speech attribution, visual salience, alignment timeline) and the `global_pass` — large; avoid when the rollup suffices. Paginate with `limit` and `cursor` (pass the returned `next_cursor` as `cursor`).