Generate AI video from Claude Code (or any MCP client) on your own keys

The short answer

How to generate AI video from Claude Code, or any MCP client, on the workspace’s own provider keys. Dated 2026-09-18.

1. Get access. Request an invitation on the developer tile, or subscribe. Checkout creates the workspace. The live door is Request an invitation.
2. Connect the client. Claude.ai signs in (no key to copy). Claude Code, Cursor, a CLI, or a script uses the public MCP URL plus a connection key. Step-by-step sign-in is on Connecting Claude.
3. Quote with dry_run. The render route with dry_run: true spends nothing and needs only read. The response includes total_usd and cap. Public skills also name a read-only estimate_render for the same job: price it before anything hits a provider.
4. Render on your keys, inside the cap. Money moves at render, on the workspace’s own provider keys, at the lab’s list rate. lotgen does not resell credits. A workspace or lot cap is the bound the agent cannot exceed.
5. Review. Work an agent made is labelled in the editor. Poll the job until it succeeds or fails, then read the take off the shot. Share a review link if a person has to sign off.

Twelve keys lotgen.ai pastes: Google, BytePlus ARK, fal, evolink, Kling, OpenAI, MiniMax, Black Forest Labs, Alibaba, ElevenLabs, xAI and Runway. Luma and Sora are models on fal or evolink, not keys. BytePlus ARK is the console that issues the Seedance key, and the name to put on the invoice. The invitation form also lists Anthropic, which buys text and nothing this page renders.

What “MCP server video generation” means here

On this page, MCP server video generation is one workspace graph, exposed twice: the web app, and the same /v1 routes as MCP tools. Claude, Cursor, or a script plans lots, quotes renders, and ships takes through the same doors the UI uses. Rendering spends on the keys you already pasted. The seat is the studio. The labs bill the seconds.

That is a different job from wrapping one lab.

Anything you can do in the app, an agent can do over MCP: the same workspace, the same quotes, the same caps. There is no tool catalog on this page. Connecting Claude is explicit: the tool list is read from the running server so the page cannot list a tool that was removed or miss one that was added, and the landing prints the live count the same way. After you connect, the tools your key is allowed to see are the list that matters.

A key starts lgk_. Four scopes, from Developers: read / write / render / admin. read is implied by the other three. Rendering spends on the workspace’s own provider keys; a write key spends nothing. A tool the key cannot call is not in the list it is offered. Reaching one anyway answers insufficient_scope naming the scope.

The REST contract, the SDK (npm i lotgen), rate-limit headers, webhooks, and OpenAPI live on Developers. This page does not copy them. It is the how-to: connect, quote, render, stay inside the cap.

The sibling workflow for briefing ad takes (not this page) is how to make UGC ads with AI.

Connect Claude Code, Claude.ai, Cursor, or a script

LOTGEN speaks MCP at https://mcp.lotgen.ai/mcp. The path matters: the bare origin answers 404. The /mcp URL without a key answers 401 (“sign in, or send a workspace key”). That is the endpoint up, not an outage.

Pick the door that matches the client.

Claude.ai (sign-in grant). Custom connector at that URL. Authentication: always required. Sign in and allow it on the consent screen. There is no key to copy and nothing to paste into a config file. The grant is per client, per person, per workspace. A second client is a second grant. The consent screen names the same four scope presets the API keys tab uses (read only, build, build and render, full access). Leave scope out and it asks for everything. The walkthrough is Connecting Claude. Do not treat this paragraph as a substitute for those three steps.

Claude Code, Cursor, CLI, cron, or your own script (connection key). A browser sign-in needs a browser. For anything that cannot sign in, mint a connection key from the MCP page once you have a workspace. Paste the public server block where the client keeps its servers. This is the mcpServers.lotgen JSON from Developers. It is a placeholder. Do not paste a real key into a repo, a chat, or this page.

{
  "mcpServers": {
    "lotgen": {
      "url": "https://mcp.lotgen.ai/mcp",
      "headers": { "Authorization": "Bearer lgk_..." }
    }
  }
}

Revoke a grant or a connection key from the MCP page. Revoking is immediate and not reversible; the lots and takes already made stay. An agent has at most your permissions and never more.

Skills, so the agent knows the order. The tools are a surface, not a procedure. Knowing a quote tool exists does not tell an agent to quote before rendering, or that creating a lot spends nothing. The public library is github.com/ascendedent/lotgen-skills: the skills that sequence the MCP tools, and the ones that teach the craft (prompt, light, motion, cost). Install as a Claude Code plugin:

/plugin marketplace add ascendedent/lotgen-skills
/plugin install lotgen@lotgen

Then run /mcp and sign in. Or clone the repo and copy the skill folders into the agent’s skill directory; you still connect MCP yourself. No credential is shipped in that repository.

A key cannot mint keys. That is a signed-in person’s job, on purpose.

Quote before you spend

The product match for an agent loop is a dollar figure in the tool response, plus a cap the loop cannot blow through. A Veo channel without a pause can wipe a budget in hours; an agent makes that worse. Do not skip the quote.

The REST quote, from Developers: the same render route with dry_run: true. It spends nothing and needs only read. The public example response shape:

{ "dry_run": true, "total_usd": 0.42, "estimate": { }, "cap": { } }

total_usd is the quote. cap is the bound. Every route that renders takes dry_run. A dry run counts as a read, not a render, for rate limits.

The skills quote. The public cost skill names estimate_render as the read-only call that prices one take on a given provider, including what that provider will change about the shot (a five-second ask that becomes six, a reference the model will drop). Creating a lot never renders and never spends. The money moves at render_scene and render_project, and nowhere else, plus post passes. If the tool list on your key does not show a name this page used, believe the list your key was offered, not this page. This page does not invent argument schemas.

Caps and headroom. Pricing: spend caps per workspace and per lot, usage by take. Read the cap (get_workspace in the public cost skill is read-only) before you render. If total_usd does not fit, do not render. Clearing or setting the cap is a write (admin / workspace settings), not something an agent should reach for to answer “how much is left.”

Rate limits, from Developers: per key, per minute, 600 requests and 30 calls to routes that render. One render call is one render however many jobs it queues. An MCP tool call is two requests, because the server confirms the key first. Treat the X-RateLimit-* headers as the truth; the window is per API replica.

The $0.42 figure above is the public docs example (and the homepage illustration of a 5s Kling 3.0 take at 720p). It is a worked example, not a measured run.

Paste your own spec into the AI video cost calculator. Toggle “I’ll run this via API or MCP” if that is the job.

Render a shot from Claude Code

Once Claude Code (or Cursor) is connected, the sequence the public contract walks is five calls. MCP tools are those same /v1 routes, filtered to the key. In prose, not as a secret schema:

1. Create a lot. create_project (REST: POST /v1/projects). default_provider is any provider the workspace holds a key for. The public example uses "fal". Creating the lot spends nothing.
2. Add a shot. add_scene with a prompt. Public example: "a bottle on a beach at dawn". Still free.
3. Quote. dry_run: true on the render route, or estimate_render if that is the tool your key lists. Read total_usd and cap. Read incompatibilities if the quote names them. Tell the person before spending.
4. Render, if it fits. render_scene without dry_run. REST answers 202 with a job to poll. Skills: this is one of the two calls that spend (the other is rendering the whole lot).
5. Review the take. Poll get_job until succeeded or failed. Then read the shot’s assets and a signed URL for the bytes. Work the agent made is labelled in the editor. A review link is how a person signs off without seeing keys or costs.

A model of the Claude Code turn, still not a log:

Quote a 5-second shot, prompt “a bottle on a beach at dawn,” on the fal key already on this workspace. If total_usd fits the cap, render it. If it does not, stop and show me the quote.

What the agent should do, from the public cost skill’s order: confirm a working key (list_providers), quote (estimate_render / dry_run), read the cap (get_workspace), then render_scene. What it should not do: skip the quote, retry a write that was refused, or invent a tool that is not in its list.

The SDK path is the same graph without MCP:

import { createLotgenClient } from "lotgen";

const api = createLotgenClient({ token: process.env.LOTGEN_API_KEY! });
const { project } = await api.createProject({ name: "Spring drop", default_provider: "fal" });
const { scene } = await api.addScene(project.id, { prompt: "a bottle on a beach at dawn" });
const quote = await api.renderScene(scene.id, { dry_run: true });
// quote.total_usd, quote.cap. Then render only if it fits.

That snippet is from Developers. Full curl, error shapes, and OpenAPI stay there.

Failed generations: the lab’s rule still applies. Kling’s developer page says failed video generations do not deduct Units. Gemini Veo: charged only if the video is successfully generated. Other labs differ. Quote the successful take. This page does not invent a lotgen keep rate.

Kling MCP vs Veo MCP

You do not install a second vendor MCP to switch labs. You pick a key on the lot.

“kling mcp” as a search is usually a GitHub wrapper around Kling’s own API. “veo mcp server” is usually the same idea for Google. On lotgen, Kling and Veo are keys the workspace already holds (Kling direct, or Kling as a model on fal/evolink; Veo on the Google key). The agent names the provider the shot should use. The quote is that provider’s list rate, on your key, in total_usd.

Dollars for an 8-second clip, both bills: Kling vs Veo cost. Veo tiers (Standard / Fast / Lite) live on Veo 3 cost per second. On that page, Veo 3.1 dollars are labeled Gemini vs Vertex. Gemini audio is the default; do not mix Vertex silent rows into a Gemini cell. Veo 3.1 lengths on the predict path are 4 / 6 / 8 only. Kling credits vs API Units: Kling pricing explained. This page does not redo those tables.

Same-prompt Kling vs Veo footage is a compare still to come. It needs a real run first. This page does not invent it.

A US lotgen workspace cannot use BytePlus ARK direct. Seedance still belongs for keys that can paste it; US readers take fal/evolink Seedance, at marketplace rates. That is a product fact, not a lab ranking.

What it costs

MCP is not a separate SKU. Pricing: MCP on every plan. Claude over MCP is the same price as the UI.

What a second costs across the twelve keys is how much AI video generation costs. Paste your own spec into the AI video cost calculator (toggle “I’ll run this via API or MCP” if that is the job). Those grids are not copied here. They refresh on their own cadence.

A dry-run quote is free. Planning a lot is free. The seat is the product. The labs are the other line on the card.

Request access

lotgen.ai is invite-only beta. The door is the invitation form. Pick Developer or agent builder (“You want the API or MCP more than the studio.”). The form does not create an account and it charges nothing. A person reads it. If it is a fit, the invitation arrives by email and creates the workspace when you accept it.

Request an invitationSeats and pricingYou want the API or MCP more than the studio.

Or start today: Solo is $99 a month and checkout creates the workspace. Same graph, same MCP, no queue. Seats and pricing. The API/MCP contract, once you have a key: Developers.

Prices checked on 2026-09-18: how to re-check yourself

Refresh this page when the MCP URL, the scopes table, or the Solo and Team cards move. Minimum: on each major model release.

1. Homepage MCP line: lotgen.ai (the live tool count; MCP at https://mcp.lotgen.ai/mcp).
2. Contract: Developers (keys lgk_, scopes read / write / render / admin, dry_run, Partner $149, the mcpServers.lotgen snippet).
3. Sign-in walkthrough: Connecting Claude (path matters; bare origin 404; tool list from the running server).
4. Seats: Pricing (MCP on every plan; Solo $99; Team $90 min 3, $79 from 6). Check whether the Partner price is on that page yet too.
5. MCP endpoint: https://mcp.lotgen.ai/mcp without a key should be 401, not 404.
6. Skills README: ascendedent/lotgen-skills.
7. Render dollars: how much AI video generation costs, AI video cost calculator, Kling vs Veo cost, Veo 3 cost per second, Kling pricing explained.

Related on this site: one AI video API across your keys · how to make UGC ads with AI · AI UGC ads cost · best AI video generator for ads · Developers · seats and pricing · request an invitation

Request an invitationSeats and pricingQuote before the agent spends.

FAQ