Luma AI Agents: Boards, Skills, MCP and What Agent Work Costs

Luma Agents is how Luma AI works today: a board where an AI collaborator plans a project with you, then generates the images, video and audio on the same canvas. It launched in March 2026, and new users start there instead of the app once called Dream Machine. Luma's pages put no price on agent time. Every generation the agent runs is paid in credits at that model's own rate, drawn from the 10,000 credits a month that come with Plus at $30 a month billed monthly or $25 a month billed yearly ($300 a year). Our verdict: it pays off for teams that brief, iterate and hand off visual work in one place. Skip it for occasional clips.

A board, an agent and the models behind it

Work in Luma happens on boards: infinite canvases where people and agents collaborate in real time, with semantic search across the assets. A new version branches out on the canvas instead of overwriting the last one. Behind the chat, the agent routes each task to a model family: Ray, Veo or Kling for video, and Uni-1, Nano Banana, Seedream or GPT Image for stills. Advanced users can pick a model by hand instead. Kling, Seedance, Veo and other outside models in Luma can change in availability, length limits and price (review).

Boards are private by default. You share one by link or email invite with Viewer or Editor permission; Viewers cannot generate or modify assets, Editors can, and Plus lists edit access for guest collaborators. A board holds up to 10,000 objects, and Pages splits a large one into sections. Teams need one more check. Luma's pricing page lists member management under its Team plan, while an April 2026 post said accounts on all plans can add members, so confirm with Luma before buying for a team.

Research Mode first, Create Mode second

The agent has two modes, and the order matters for your bill. Research Mode asks targeted questions, gathers references and analyzes what is already on the board, then writes editable plans: creative briefs, campaign outlines, shot lists, storyboard structures and production steps. It generates no media, and the gathered context carries forward into production. ^[1] Luma's learning center also advises starting there because planning costs less than generating. One learning-center page calls the same mode Brainstorm Mode.

Create Mode is where credits move. It generates images, video and audio, animates stills, extends scenes, adds voiceovers, lip sync and captions, and reframes footage for each platform (video-to-video guide). You can feed it PDFs, scripts, images, videos, decks and audio. Two habits keep output on track. First, keep pointing the agent back at your master reference assets, because Luma warns it can drift and hallucinate key elements between versions. Second, ask for one version when one is enough.

That second habit is about money. By default the agent may generate several variations or retry outputs on its own, which can spend credits silently unless you ask for a single result. ^[2] Failed generations get their credits back, but takes you reject still cost credits (pricing). Luma also suggests drafting at low resolution and upscaling a good draft instead of regenerating it.

Skills and mini apps: repeat what worked

Skills arrived in June 2026. A Skill is a reusable workflow on a board, built from an input, the tools and plain-language instructions. You create one from a prompt, a finished output, a board process or a workflow document, or by forking an existing Skill; built-in and shared Skills are read-only until forked. A Skill runs from the Skills menu, by slash command, by typing @ in the chat, or when the agent chooses it. A magic link imports it into another board. That is the point. One person's recipe becomes the studio's routine, as our agency guide shows for client work.

Mini apps are packaged workflows in the Solutions tab: Face Swap, Camera Angles, Light Transfer, Relight Studio, Virtual Try-On, Background Swap, Ad Variants and Product Shots. Each one plans, generates, checks its own output and returns finished assets. Product Shots returns three images per run and keeps the product's shape, label and packaging while changing the scene. Background Swap takes one to eight images per run. Face Swap handles images only.

Scenes, Layers and Ad Variants

Scenes (August 2026) plans a storyboard of keyframes and renders nothing until you approve it, then renders with Ray3.2 or Seedance 2 in 9:16 or 16:9. It is capped at roughly 10 frames including references, so it suits short pieces. The Ray3.2 path is silent. There is no propagate-to-all edit, so changes across keyframes happen frame by frame, and Luma says large automated variant batches are not yet a workflow to rely on.

Layers (July 2026), powered by Uni-1, splits an image into editable elements on transparent backgrounds, such as a product, a headline or a background. It works on still images, not video, and modifying a layer keeps the original hidden in a group. The pricing page charges 75 credits per image at 1K, 105 at 2K and 175 at 4K. Ad Variants (October 2026) rebuilds one approved static ad in five placements, from Story 9:16 to Widescreen 16:9, and in English, German, Italian, Japanese or Korean. Its first release covers resizing and translation only.

Connectors and the MCP for Claude and ChatGPT

The agent reaches five services: Google Drive, Dropbox, Frame.io, Box and Airtable. They are read-and-add only. Luma can read, import, create, upload and save, but never deletes or renames anything in the connected account; the Airtable connector can also update records. Sign-in uses OAuth, so Luma never sees the password, and every connector stays off until a team admin enables it. Google Drive shows Luma only the files you pick; Frame.io lets it read review comments with exact timecodes.

In November 2026 Luma released an MCP connector for Claude and ChatGPT. It uses your normal Luma sign-in, needs no API key and no separate subscription, and bills the Luma team you choose at sign-in, at app credit rates. ^[3] A team without credits will not run generations. In Claude you add it by URL, and on Claude Team or Enterprise an admin adds it once for everyone. Results never land on a board by themselves: save them to the Asset Library first. Models offered in ChatGPT may differ from the app's. Luma Agents themselves are stated for the browser workspace; the iOS app syncs your account and credits, as our app guide explains.

What agent work costs on each plan

Nothing on Luma's pages prices agent time. Credits pay for output. Plus costs $30 a month billed monthly or $25 a month billed yearly ($300 a year), with 10,000 credits a month. Pro costs $90 a month or $75 a month billed yearly ($900 a year), with 40,000 credits. Ultra costs $300 a month or $250 a month billed yearly ($3,000 a year), with 150,000 credits. ^[4] Team and Enterprise are priced by sales.

Pro lists "4x usage with the Luma Agents" and Ultra "15x", and Luma does not define the term. Separately, by our arithmetic, Pro's and Ultra's credit allowances are exactly 4x and 15x the Plus allowance. For scale, a 5-second 1080p SDR Ray3.2 clip costs 400 credits at the table price, so Plus covers 25 of them by our arithmetic, if every credit goes to that one job with no failed or rejected takes. Unused credits expire at the end of each billing cycle. Extra credits bill after use under an optional spending limit, with no per-credit price on the pricing page, and used ones are not refunded.

The bottom line

Luma Agents turns Luma AI into a shared production board: plan in Research Mode, generate in Create Mode, then repeat what worked through Skills and mini apps, or drive it from Claude and ChatGPT. Luma's pages put no price on agent time; every generation the agent runs is paid in credits at that model's rate, and it may add variations unless told otherwise. That suits teams with a steady flow of briefs and approvals. For the odd clip, a monthly allowance that expires each cycle is hard to justify.