each::labs

Tạo và chỉnh sửa hình ảnh, video, nhạc và giọng nói bằng các mô hình each::labs, đồng thời chạy các quy trình làm việc each::labs của tổ chức bạn từ bất kỳ ứng dụng AI nào.

Máy chủ MCP được lưu trữ

npx add-mcp 'https://mcp.eachlabs.ai/mcp'

Cài vào Claude Code, Codex, Cursor và nhiều công cụ khác

Tài liệu

MCP Server

Use each::labs models, workflows and the Video API from any app that supports MCP, through the hosted server at mcp.eachlabs.ai. Sign in with your each::labs account; no API key.

MCP (Model Context Protocol) is an open standard that lets AI apps use outside tools. The each::labs MCP server is hosted at:

https://mcp.eachlabs.ai/mcp

Connect

Add https://mcp.eachlabs.ai/mcp as a remote MCP server, with the Streamable HTTP transport, in any app that supports MCP. The app opens the each::labs sign-in page, you choose your organization, and every run is billed to that organization. There is no API key to create or paste.

Apps with a chat interface can show interactive cards. Apps without cards, such as terminals and editors that do not support them, return the same results as text and ask you in the chat for file links or values.

Using a coding agent? Paste this to it: Fetch and follow https://docs.eachlabs.ai/agent-setup.md to connect me to each::labs. The Agent setup page has exact steps for Claude Code, Codex CLI, Cursor, VS Code, OpenCode and Claude Desktop.

Searching these docs from your editor is a separate, read-only server: Documentation as MCP. It does not generate media.

Using Claude? Use each::labs in Claude has the steps for claude.ai, Claude Desktop and Claude Code.

Using Claude or Codex, you can also add the each::labs plugin: the same connector plus skills for common media tasks.

Environments without a browser (CI, servers) can run the video tools locally with an API key instead. See Headless and CI.

Sign-in and billing

The first call opens the each::labs consent page in your browser. If you are not signed in to the dashboard, the page sends you through sign-in and back.

On the consent page you:

  1. Choose the Organization. The page states it plainly: Predictions will run and bill as <organization>.
  2. Review the permissions the client asks for (the scopes below).
  3. Click Approve.
WhatValue
CredentialAn OAuth token that your client stores. It never appears in the conversation and is never shown to the model.
Billed organizationThe one chosen on the consent page. Every job behaves as if that organization had called the API itself, at the same prices as the REST API.
Signing in againThe token is refreshed in the background while you use it. You sign in again after 30 days without use, or 90 days after you signed in, whichever comes first. To bill a different organization, disconnect the server in your client and connect again.
If sign-in failsIf the consent page shows organization has no runner credential, open the dashboard once as that organization and try again.

Scopes

A scope is a permission your client asks for at sign-in. The consent page lists each one. A tool that needs a scope your token lacks is refused before anything is submitted, and nothing is billed for a refused call. A connection approved before a scope existed does not have it; disconnect and connect again to approve the full set.

ScopeGrants
predictions:createRun models and video tools, check on their jobs, and list the organization's recent runs
predictions:readRead prediction status and results, and list the organization's recent runs
workflows:runTrigger workflow executions, and list and read workflows and executions
workflows:readList and read workflows and executions
uploads:createShow cards in the conversation, and upload the files you give on the upload card

The model catalog tools (search_models, get_model_schema) need no scope.

Long jobs

A tool that starts a job waits up to about 20 seconds for it. A job that finishes in that time returns its result in the same call. A longer job returns its id with status: "running" within those 20 seconds, and the AI keeps checking with that id until the job finishes. The job keeps running on each::labs the whole time. Returning early keeps your app from giving up on a call that waits too long.

The call that returned runningContinue with
run_model or a video toolget_job with the prediction_id
run_model with count above 1get_job with every id as prediction_ids; the variants show together once all finish
trigger_workflowget_workflow_execution with the execution_id

Billing happens once, when the job runs, not per check. Long videos can take 5 to 10 minutes. The AI keeps checking until yours is ready, and asks you before trying a different model.

Recent runs

list_recent_jobs shows your organization's most recent runs, newest first: id, model, status, time and, once a run is done, its output link. The AI uses it to pick up a run it started earlier, for example after a brief connection interruption, and keeps checking that run with get_job rather than starting a new one. The list never includes prompts or inputs, and it shows the same runs as the runs list in your dashboard.

Tools

Tool names are lowercase with underscores. A call with a parameter the tool does not have is refused before anything runs, and nothing is billed. Each tool also tells your app whether it only reads or starts a billed job; your app uses this to decide which calls need your confirmation. The full list with parameters, defaults and return values is the MCP tool reference.

Video

One tool per Video API capability, plus run_ffmpeg and get_job: 46 tools on the hosted server. Each takes media by URL and returns the result by URL, except the four that measure or detect, which return numbers. With the eight model and workflow tools and the 13 card tools below, the hosted server offers 67 tools. Per-client setup is in the video MCP quickstart.

ToolDoesScope
trimCuts a clip by start time and length, without re-encodingpredictions:create
segmentSplits a video into equal-length partspredictions:create
concatJoins 2 to 10 clips into one videopredictions:create
concat_copyJoins 2 to 50 parts that share one format, without re-encodingpredictions:create
crossfadeJoins two clips with a crossfade between thempredictions:create
speedChanges playback speed, from 0.25x to 4xpredictions:create
reversePlays a video and its audio backwardspredictions:create
cropCrops a region of the framepredictions:create
scaleResizes a video, keeping its proportions if you askpredictions:create
padPlaces a video on a larger canvas with a background colourpredictions:create
rotateRotates a video by 90, 180 or 270 degreespredictions:create
reframeChanges the aspect ratio for social formats (9:16, 1:1, 4:5, 16:9) by cropping or adding barspredictions:create
reframe_trackReframes to a smaller frame that follows a subject over timepredictions:create
colorAdjusts brightness, contrast, saturation and gammapredictions:create
lut3dColour-grades a video with a .cube LUT filepredictions:create
fadeFades a video and its audio in, out or bothpredictions:create
overlayPlaces a second video over a base video at a corner or the centrepredictions:create
watermarkPlaces an image or video over a video as a watermarkpredictions:create
sticker_overlayShows a sticker or logo over a video during a time windowpredictions:create
title_cardDraws a text title over a videopredictions:create
captionsBurns subtitles (SRT, VTT or ASS) into a videopredictions:create
subtitle_composeCreates animated, word-by-word captions from word timings and burns them inpredictions:create
extract_audioExtracts the audio track as mp3, wav or m4apredictions:create
audio_replaceReplaces a video's audio track, or mixes in a second onepredictions:create
audio_duckLowers background music automatically while a voice speakspredictions:create
audio_masterMasters audio to a loudness target for publishingpredictions:create
loudnormNormalises loudness to a target levelpredictions:create
volumeChanges the volumepredictions:create
audio_analysisMeasures loudness, loudness range and peaks; returns numbers, not a filepredictions:create
silence_detectFinds silent parts; returns their times, not a filepredictions:create
silence_removeKeeps only the time ranges you give and joins them into one filepredictions:create
silence_splitSplits a file at its silent partspredictions:create
scene_detectFinds scene cuts; returns their times, not a filepredictions:create
stillsTakes one frame as a jpg or png imagepredictions:create
thumbnailsTakes frames at a regular intervalpredictions:create
posterReturns the video plus a poster imagepredictions:create
storyboard_spritesPacks frames into one sprite sheet plus a thumbnail index for video playerspredictions:create
gifMakes an animated GIF from a clippredictions:create
slideshowTurns 2 to 50 images into a videopredictions:create
transcodeConverts a video to another format, codec or qualitypredictions:create
hls_ladderEncodes several quality levels for HLS streamingpredictions:create
package_abrPackages a video for streaming as HLS and DASHpredictions:create
package_abr_ladderPackages several quality levels for streaming as HLS and DASHpredictions:create
probeReads a file's details: format, streams, length, resolution; returns numbers, not a filepredictions:create
run_ffmpegRuns an ffmpeg command on the input URLspredictions:create
get_jobReturns the status and result of a prediction, or of several variants from one run_model callpredictions:read or predictions:create

Model catalog and any model

Three tools cover the whole model catalog, not only video, and list_recent_jobs lists recent runs. The three use the same catalog and the same prediction as List models and Create prediction.

ToolDoesScope
search_modelsFinds models by name; returns slug, name, type and pricing for each matchnone
get_model_schemaReturns a model's input schema, so an input can be built from itnone
run_modelRuns a model by slug with an input matching its schema; returns the result, or the job id while it is still running. With count (1 to 8) it runs that many predictions of the same input and returns them togetherpredictions:create
list_recent_jobsLists the organization's most recent runs, newest first: id, model, status, time and output link. Used to pick up a run the AI started earlierpredictions:read or predictions:create

Workflows

Run and read the workflows your organization built in the dashboard. Creating or editing a workflow stays in the dashboard.

ToolDoesScope
list_workflowsLists the organization's workflows: id, slug, name, status and countsworkflows:read or workflows:run
get_workflowReturns one workflow: its definition, versions and input schemaworkflows:read or workflows:run
trigger_workflowTriggers a workflow version with inputs; returns the outputs, or the execution id while it is still runningworkflows:run
get_workflow_executionReturns an execution's status and outputsworkflows:read or workflows:run

These tools do not set up webhooks; the AI checks on a run itself until it finishes.

Cards

Tool inputs are URLs. A file attached to a conversation cannot reach a connector tool, so the server provides small interactive cards that your app shows inside the conversation: one to upload files, and others to pick settings, choose a variant, fill in a workflow's inputs or choose a range or crop. Each card has a matching wait tool that the AI calls right after showing it. When you press the card's button, the AI receives your choice and continues; you do not need to type anything.

ToolDoesScope
request_fileShows the upload card with a one-line prompt and the kinds it accepts (image, video, audio). count is the number of files expected; max is the most the target model or tool takes. A card for several files takes up to max, or up to 50uploads:create
wait_for_uploadWaits for the files the person uploads on the card and returns their public URLsnone
presign_uploadCalled by the upload card, not by the AI or by you: prepares the upload of the file you pickeduploads:create
file_uploadedCalled by the card after the upload: returns the file's public_urlnone
request_parametersShows a model's settings as a form, built from its input schema, with a Run buttonuploads:create
wait_for_parametersWaits for Run and returns the model, version and input to pass to run_modelnone
request_workflow_inputsShows a workflow version's inputs as a form, with a file picker for each image, video or audio inputuploads:create
wait_for_workflow_inputsWaits for Run and returns the workflow, version and inputs to pass to trigger_workflownone
request_trimShows a video or audio file with a timeline and two handlesuploads:create
request_cropShows a video's first frame or an image with a rectangle to move and resizeuploads:create
wait_for_editWaits for the person's range or rectangle and returns the values to pass to trim or cropnone
wait_for_choiceAfter a run that asked for several results, waits for the variant the person submits and returns itnone
show_filesShows files the AI already has a URL for, such as an earlier result, on a result cardnone

Every tool that returns media also carries the result card: the image, a video or audio player and download buttons. When a run asked for several results, the result card is a picker: the person selects a variant, as often as they like, and presses the button under the variants once; it reads Use variant and the number selected. A workflow's outputs show together as a gallery. An edit of a file the person gave gets a Compare button for a before-and-after view.

One upload card takes up to 50 files of up to 100 MB each and uploads four at a time; too many files or a file over 100 MB is refused before anything uploads. The workflow form's file fields follow the same limits. A card used once keeps showing that choice when the conversation is reopened.

Cards are MCP Apps served by this server, so they show in any app that supports MCP Apps. In apps without card support the AI asks in the chat instead, and results come back as text. A card that cannot load in your app says to use the chat instead. How to use the cards is on Use each::labs in Claude.

This documentation is built and hosted on Mintlify, a developer documentation platform.