v2.6.0 — Translate video, project editing, and video edits
Translate the speech in a video into several languages in one run with POST /v1/workflows/translate-video, or dub a single file with POST /v1/tools/dub-video. Each language becomes its own editable video, and export and video edits can target one video in a multi-video project with editableVideoId. This release also adds project editing: inspect a project’s videos and saved workflows, then apply typed edits with safe retries and undo. Remix actions are now called video edits: workflow starts take videoEdits, and POST /v1/projects/{projectId}/video-edits replaces the remix endpoints. MCP clients can delegate work to the VideoGen assistant and import video links directly.
Added
- Translate video workflow:
POST /v1/workflows/translate-videotakes an uploadedvideoFileId(up to 8 minutes) andoutputLanguages(one or more language codes, such asesandfr). Repeated languages are translated once. PasssourceLanguagewhen you know the spoken language; otherwise it is detected. The original speakers’ voices and timing are kept. Languages are built in parallel, and one failing language does not stop the others. The run fails only when every language fails. Optional fields arevideoEdits(applied to every language’s video),autoExport(renders the first successful language),exportOptions, andisOutputTemporary. Credit use grows with the video’s length and the number of languages. Requires a Pro subscription. - Workflow run outputs:
GET /v1/workflows/runs/{workflowRunId}now returnsoutputs, one entry per editable video the run produced, with itseditableVideoId(vg_vide_...),language, andstatus(succeededorfailed). Translate-video runs list one entry per requested language. - Translate a video’s speech:
POST /v1/tools/dub-videotakesvideoFileIdandtargetLanguage, plus an optionalsourceLanguage, and returns one translated video file without creating a project. Billed per second of output video. Requires a Pro subscription. - Dub videos when translating a project: the
TRANSLATE_PROJECTvideo edit acceptsdubVideos(defaultfalse). Whentrue, the speech in the project’s uploaded videos is also dubbed into the target language, billed per second of each video with speech. Whenfalse, those videos keep their original audio and their captions are still translated. Returns 400 if the target language does not support dubbing. - Video edits endpoints:
POST /v1/projects/{projectId}/video-editsapplies an ordered list ofvideoEditsto one video, named by the requirededitableVideoId. SetsaveAsNewProjectto edit that video in a copy. The response returns onevideoEditId(vg_vedt_...) per edit.GET /v1/projects/{projectId}/video-editslists their status and progress, optionally filtered byeditableVideoId. See Video edits. - Project artifacts:
GET /v1/projects/{projectId}/artifactslists a project’s artifacts with theirartifactId(vg_arti_...),type,name,revision, and the matchingeditableVideoIdorworkflowId.GET /v1/projects/{projectId}/artifacts/{artifactId}reads one artifact’s saved workflow settings or video identity. Page long text withtextOffset/textLimit(UTF-16 code units) and storyboard scenes or slides withitemOffset/itemLimit. Reading never renders or generates anything. - Editing capabilities:
GET /v1/projects/{projectId}/editing-capabilitiesreturns whether you can edit the project (canEdit), every supported commandtypewith itstargetandexecution, and requestlimits(maxCommands,maxBodyBytes). - Inspect a video:
GET /v1/projects/{projectId}/videos/{editableVideoId}/inspectionpages through a video’s timeline. With no selector it returns section summaries. PasssectionIdfor a section’s layers,layerIdfor a layer’s assets, orassetIdfor an asset’s transcript words and source timings. Page withoffset/limit, and pass the returnedrevisionback asexpectedRevisionso a change between pages returns a conflict instead of mixing versions. - Edit a video:
POST /v1/projects/{projectId}/videos/{editableVideoId}/editsapplies up to 100 typed commands in one atomic batch (mode: "COMMANDS",expectedRevision,mutationId,commands). Commands cover text, captions, sections, layers, asset position, trim, playback, filters, transitions, voiceover text, and background music. These edits do not generate media or export. - Background video edits: the same endpoint and the MCP
edit_videotool also accept background commands on one video:SET_LOGO,REMOVE_LOGO,APPLY_TRANSITIONS,ADD_ZOOM,CLEAN_UP_TRANSCRIPT,CONVERT_IMAGES_TO_VIDEOS,REGENERATE_IMAGES,UPSCALE_ASSETS,CHANGE_NARRATOR,SHUFFLE_STOCK_VISUALS,GENERATE_MUSIC,TRANSLATE_PROJECT,ADD_MEDIA,ADD_DRAFT_MEDIA,GENERATE_DRAFT_MEDIA, andRUN_TOOL_ON_ASSET. Send a background command as the only command in its request, usingsectionIdandassetIdvalues from the video’s inspection. The receipt returnsoutcome: "STARTED"and avideoEditId. Follow it withGET /v1/projects/{projectId}/video-edits. Retrying the same request returns the samevideoEditIdand does not start a second task. Generation commands spend credits. Editing capabilities list these commands withexecution: "BACKGROUND". - Edit project documents:
POST /v1/projects/{projectId}/mutationsatomically changes the project title and description, artifact names, and saved workflow settings (script, storyboard scenes, slideshow slides, captions, aspect ratio, quality, voice speed, and background music). List the inspectedrevisionof each changed document inexpectedDocuments. Workflow settings changes do not regenerate or alter videos that were already generated. - Safe retries and undo: every edit takes a new caller-generated UUID
mutationIdand returns a receipt withoutcome(APPLIED,NO_OP,CONFLICT, orSTARTED) andprojectRevision. A stale revision conflicts instead of overwriting newer changes. Retry the same request unchanged to get the original receipt, or read it withGET /v1/projects/{projectId}/mutations/{mutationId}.POST /v1/projects/{projectId}/mutations/{mutationId}/undoapplies the inverse as a new edit (send a freshmutationIdin the body) and keeps other edits made since. Undo the resulting edit to redo. Undo never repeats generation or billing. - Access to projects created in the app: API keys and OAuth integrations can read and edit app-created and API-created projects that their authenticated caller can access on the selected team. Existing team and project permissions apply. Deleting an API key revokes its access.
- Choose a video from a multi-video link:
POST /v1/files/import-urlaccepts an optionalvideoIndex(1 to 20). When a link contains more than one video and novideoIndexwas sent,GET /v1/files/imports/{fileImportId}ends asfailedwith aselectionobject listing each video’svideoIndex,title, anddurationSeconds. No video is downloaded. Ask the user which video they want, then start a new import with the samesourceUrland theirvideoIndex. Do not pick the first video on the user’s behalf. - SDKs and CLI:
- TypeScript
workflows.translateVideo/translateVideoAndWait,tools.dubVideo/dubVideoAndWait, andprojects.startVideoEdits/startVideoEditsAndWait/listVideoEdits. - TypeScript project editing:
projects.listProjectArtifacts,getProjectArtifact,getProjectEditingCapabilities,inspectEditableVideo,editVideo,applyProjectEdits,getProjectEdit, andundoProjectEdit(pass the undo’s own id asnewMutationId). - Python has the snake_case equivalents (
translate_video,dub_video,start_video_edits,list_video_edits,undo_project_editwithnew_mutation_id, and so on). - The CLI adds
videogen workflows translate-video,videogen tools dub-video,videogen projects start-video-editsandlist-video-edits, and matching project editing commands such asedit-videoandundo-project-edit.
- TypeScript
- MCP:
translate_videoanddub_video.export_projectacceptseditableVideoId.start_video_editsandlist_video_editsreplaceremix_projectandlist_project_remix_actions.- Project editing:
list_project_artifacts,get_project_artifact,get_project_editing_capabilities,inspect_video,edit_video,edit_project,edit_workflow,get_project_edit, andundo_project_edit. - Assistant:
start_assistant_chathands a new project request to the VideoGen assistant, andsend_assistant_messagecontinues work on an existing API-created project.get_assistantreads the conversation,get_assistant_messagepolls a reply, andact_on_assistant_actionapproves or rejects an action the assistant proposed. - Video link imports:
import_video_from_urlstarts a link import andget_video_url_importpolls it. A successful import returns afileIdto pass tovideo_to_video. - Media previews: on the hosted MCP server,
show_media_previewdisplays up to 12 finished images, videos, or audio files in the conversation from theirvg_file_...ids. It does not generate or change anything.
- Agent frameworks:
translate_videoanddub_videotools in the OpenClaw plugin and the OpenAI Agents package.
Changed
- Remix actions are now video edits: workflow start requests take
videoEditsinstead ofremixActions, and workflow start responses returnvideoEditIds(vg_vedt_...) instead ofremixActionIds. The edit types and their fields are unchanged. Rename the field in your requests and readvideoEditIdsfrom responses when you upgrade. The SDK types use the new names. - Export one video:
POST /v1/projects/{projectId}/exportaccepts an optionaleditableVideoIdfrom a run’soutputs. - Project status reports stopped generations:
statusonGET /v1/projects/{projectId}andGET /v1/projectscan now befailedorcancelledwhen an unfinished project’s latest generation stopped.generatingmeans work is still pending or running. Handle the two new values wherever you switch onstatus. - Assistant action labels: on assistant messages,
actions[].labelis the approval card’s title (defaultApply edit) andactions[].detail.summaryis its description.
Deprecated
- Remix endpoints:
POST /v1/projects/{projectId}/remixandGET /v1/projects/{projectId}/remix-actions, the SDK methodsprojects.remixProject,remixAndWait, andlistProjectRemixActions(Pythonremix_project,remix_and_wait,list_project_remix_actions), the CLIremix-projectandlist-project-remix-actionscommands, and the MCPremix_projectandlist_project_remix_actionstools still work but are no longer documented or listed. Move toPOST /v1/projects/{projectId}/video-editsandGET /v1/projects/{projectId}/video-edits, which require aneditableVideoId.
Fixed
- Wrong file type fails up front: tool requests that pass a file of the wrong media type, such as a video id in
startFrameFileIdorimageFileIds, now returninvalid_parametersimmediately with the parameter, the expected type, and the file’s actual type. Previously the task started and then failed. - Short reference audio:
generate-video-clipwith a reference audio file shorter than 2 seconds now fails quickly with a message naming the file’s length, without retries or charges.