Which MCP server lets Claude Code or Cursor generate video and images?

Several MCP servers let Claude Code or Cursor generate videos and images: Sume (https://mcp.sume.com/mcp), fal (https://mcp.fal.ai/mcp), Replicate (a hosted server or a local npx server), Runway (https://mcp.runwayml.com/mcp), and Higgsfield (https://mcp.higgsfield.ai/mcp). Kling's API docs describe an agent Skill for Claude Code and Cursor, which is not an MCP server. Which one fits depends on four things: the models you need, how the server signs you in (OAuth or an API key), how each generation is billed, and whether the agent must also edit and assemble what it makes.

The table below sets the options side by side from each company's own page, read on 2026-09-29, followed by setup for Claude Code, Cursor, and other MCP clients, then Sume's own tools and billing. Sume's basics page says hosted MCP still works but is not the primary path today: it fits clients that speak remote MCP, while a backend calls the HTTP API.

Which MCP servers generate video and images in Claude Code or Cursor?

The five MCP servers add to Claude Code and Cursor the same way: a URL (or a local command) plus a sign-in. Kling's Skill installs from its own link instead. They differ in catalog, sign-in, and who bills you.

• fal and Runway each run a second MCP server that does not generate: fal's docs server at https://fal.ai/docs/mcp only searches fal's public site, and Runway Dev at https://dev.runwayml.com/mcp looks up tasks, manages model routers, and reads docs.
• Each of these has its own setup post: fal, Replicate, Runway, Higgsfield, and Kling.

How do I add a media MCP server to Claude Code?

Claude Code adds any remote server with claude mcp add --transport http <name> <url>; a server that takes an API key gets it with --header "Authorization: Bearer …" (fal's page shows exactly that). For an OAuth server such as Sume's, claude mcp login runs the sign-in from your shell. claude mcp list then shows each server's health, such as Connected, Needs authentication, or Failed to connect.

claude mcp add --transport http sume https://mcp.sume.com/mcp
claude mcp login sume
claude mcp list

How do I add one to Cursor?

Put the server in ~/.cursor/mcp.json for every project, or in .cursor/mcp.json for one project. A remote entry is a url, plus headers when the server wants an API key; Cursor supports OAuth for servers that require it. Cursor asks for approval before it uses an MCP tool by default, so the first paid call waits for your click.

{
  "mcpServers": {
    "sume": {
      "url": "https://mcp.sume.com/mcp"
    }
  }
}

How do I connect another MCP client to Sume?

Point a streamable HTTP MCP server entry at https://mcp.sume.com/mcp and run that client's OAuth login for the server name you configured. The client discovers Sume's protected-resource metadata from the endpoint and sends you to the consent page on mcp.sume.com. To check the connection, ask the agent to call tools_list (every tool the session can see) or mcp_health (endpoint, auth source, and safety posture).

Which Sume MCP tools generate video and images?

Paid tools that make or change images and video. Omit payload.model on the two generators to let Sume route to sume/auto, or pick a catalog id from image-models_list or video-router_models. Sume Image 1.0 and Video 1.0 are REST-only and not in tools_list.

Why can't the agent generate anything after I sign in?

Sume's hosted OAuth is read-only until you turn Write on. The consent page shows Read locked on and a Write toggle that is off by default; paid tools such as generate_image return insufficient_scope on a read-only session. Turn Write on to also grant mcp:write, or connect with an API key, sent as Authorization: Bearer $SUME_API_KEY or x-api-key. There is no separate paid scope. MCP server API key vs OAuth covers the trade-off.

How much does it cost to generate over MCP?

Where a server's page says how it bills, the MCP connection itself is not what you pay for; the generations are. fal says its server is free and model runs bill at the same pricing as direct API calls, and its get_pricing tool returns a model's cost before you run it. Runway's MCP uses your Runway credits, just like the Runway app. Kling's Skill notes that fees are incurred for each submission.

On Sume, each job is billed from the workspace's USD balance: one wallet. The ledger reserves the estimate before the job runs, captures it after successful completion, and releases it after a failure or cancellation before capture. The Video Router row on the API pricing page reads "Pass-through per model": provider list rates × Sume margin, with per-model rates on GET /v1/videos/models. The same page lists Timeline render at $0.10 per output minute. How much an AI video costs works through examples.

The wallet is the spend gate, so the controls are per call. Paid tools require an idempotency_key, so a retried call cannot submit the same job twice. dry_run=true previews admission and cost without submitting, and max_spend_usd caps a call only when you provide it. The Safe automation guide covers how agents should treat these boundaries.

Will a long video job time out in Claude Code or Cursor?

Video takes minutes, so these servers hand back a job and let the agent poll rather than hold one call open. Kling's Skill notes say video generation typically takes 1–5 minutes and images about 20–60 seconds. fal's run_model waits up to 45 seconds by default; for video, fal's page says to use submit_job, check_job, and get_job_result from the start, because another run_model call creates a new billable job.

Sume's generate_video answers with a job id in milliseconds, never a clip URL. The agent then calls jobs_wait, which holds at most 55 seconds per call on remote MCP, and repeats it with the same id until the job is terminal (Jobs and results). Claude Code's per-request timer for an HTTP server is at least 60 seconds, so a 55-second wait fits inside it. Waiting out long video jobs covers that loop.

Can the agent use an image from my laptop?

Not directly on Sume: hosted MCP cannot read files from your laptop. Generation inputs are URLs, and media tools such as timeline_create and avatar-image-to-video_create take Sume-hosted media, such as an earlier job's output. fal's upload_file tool takes a public file URL or base64 data and puts it on fal's CDN. MCP server file upload explains how a local file reaches a remote tool.

Which one should I pick?

Match the server to the job, dated 2026-09-29:

• Pick fal if you want the agent to search a large catalog (1,000+ models, fal says) and check a model's price with get_pricing before each run.
• Pick Replicate if you want to run the server locally with npx, or already build on Replicate's model catalog.
• Pick Runway if you already pay for Runway credits or want Runway's Gen-4.5 or Veo 3.1 from the agent; Veo is not in Sume's catalog.
• Pick Higgsfield if you want its preset workflows (paid ads, motion design, UGC) inside Claude.
• Pick Kling's Skill if you only need Kling models and your agent can run a Node.js skill.
• Pick Sume if the agent must generate and then finish the video in one server: stills and clips, lip sync, upscales, cutouts, and a Timeline cut, billed from one wallet with dry_run cost previews. If you would rather send one brief and get an edited video back, see a video agent API that turns a brief into a finished video.
• Skip Sume's MCP if you need Veo, or if your integration is a backend job rather than an MCP client: call the HTTP API instead.