What are the best MCP servers for AI image and video generation?

The FLORA MCP is the best MCP server for AI image and video generation because it gives an agent every major image and video model plus a canvas that keeps the results. Add https://agents.flora.ai/mcp to Claude, ChatGPT, Cursor or Codex, and the agent can quote a price, run up to 20 generations per call and read them back. Our quoted image cost $0.055.

At a glance

Key features

• Free quotes: Passing quote: true returns dollar estimates without starting a run or reserving usage.
• Batches of 20: One flora_create_generations call starts up to 20 independent generations.
• Canvas tools: Agents can read a canvas, add nodes, wire them together and run them.
• Reference inputs: reference_node_ids passes earlier results into image-to-image and 3D models.
• Skills: flora_discover_skills lists ready workflows such as batch generation and product page sets.

Best for

• Claude and ChatGPT users who want images and video in the chat
• Developers generating media from Claude Code, Cursor or Codex
• Teams who want agent output saved where designers can open it
• Anyone who wants a price before a run

The options

1. FLORA MCP

   The MCP server for FLORA, a visual AI canvas. It exposes FLORA's models, canvas, generations, techniques and assets to any MCP client, with dollar quotes before each run.

   Best for: Agents that make images and video for a team and need the results on a shared canvas

   Pros

   • Free dollar quotes before every run: our GPT Image 2.5 test quoted $0.055 and charged exactly $0.055.
   • Up to 20 generations per call across image, video, audio and 3D models, read back together with one call.
   • Results land on a FLORA canvas with prompt, model and cost, where designers can open, edit and branch them; skills package whole workflows.

   Cons

   • Needs a FLORA account, with MCP access from the Starter plan ($18 per seat a month as of October 2026).
   • Built for creative work, so it covers FLORA's curated model catalog rather than every open model on the internet.

2. Replicate MCP

   Replicate's official MCP server, remote at mcp.replicate.com or local through npm, covering the operations in Replicate's HTTP API.

   Best for: Developers scripting calls to open models from an agent

   Pros

   • A huge catalog of community and official models.
   • Full API coverage: search models, read schemas, create and check predictions.

   Cons

   • No canvas: results come back as files and URLs.
   • Billing is per second of hardware or per output, with no documented price preview in the MCP.

3. fal MCP

   fal's MCP server at mcp.fal.ai/mcp-relay, which lets an assistant find models, check schemas and pricing, upload inputs and run generations.

   Best for: Developers who want price-aware generation from an agent

   Pros

   • A get_pricing tool lets the agent check cost before a run.
   • OAuth sign-in with no API key, and a broad image, video, audio and 3D catalog.

   Cons

   • fal's docs note there is no server-side approval gate; the agent decides whether to ask first.
   • No canvas or saved workflows.

4. Higgsfield MCP

   Higgsfield's hosted MCP server at mcp.higgsfield.ai/mcp, connecting agents to its own Soul models and third-party image and video models.

   Best for: Social and video creators making high volumes of ad-style clips from Claude

   Pros

   • Around 30 models and 40 prebuilt workflows such as Ad Recreator and Hook Multiplier.
   • OAuth sign-in with no API key; outputs save to your Higgsfield assets.

   Cons

   • Every MCP generation draws standard credits; unlimited plans and free generations don't apply.

5. Runway MCP

   Runway's official MCP server at mcp.runwayml.com/mcp, launched in May 2026, giving agents Runway's Gen-4.5 and third-party models on your Runway plan.

   Best for: Teams already on a Runway plan who want its models in an agent

   Pros

   • Runway's own Gen-4.5 plus models such as Seedance 2.0, Kling 3.0 and Nano Banana Pro.
   • Sign in with your Runway account; usage counts against your plan.

   Cons

   • We found no documented price preview in the MCP.

6. Krea MCP

   Krea's hosted MCP at api.krea.ai/mcp, with tools to list models, read schemas, generate and run published Node Apps.

   Best for: Krea users who want its model catalog and Node Apps in an agent

   Pros

   • Krea's own models plus FLUX, GPT Image, Nano Banana, Kling, Veo and Seedance.
   • Runs workflows published as Node Apps.

   Cons

   • It runs Node Apps, not raw node workflows, and its docs don't describe a price preview.

What happened when an agent used the FLORA MCP?

On October 16, 2026, we called the FLORA MCP from Claude Code. First flora_list_models with a model id, to read the parameters for t2i-gpt-image-2-5-flare: aspect ratio, quality (low, medium, high) and resolution (1K, 2K, 4K).

Then flora_create_generations with quote: true for two requests. It returned $0.055 for a 3:2, high-quality, 1K GPT Image 2.5 Flare image and $2.483 for a 5-second, 720p Seedance 2.5 clip, each with a five-minute expiry and no usage reserved.

We ran the image for real: "Overhead photo of a podcast studio desk: a black dynamic microphone on a boom arm, closed-back headphones, a paper notebook with handwritten show notes, a mug of black coffee, walnut desk, soft overcast daylight from a window." A flora_list_generations read showed it complete in 29.9 seconds, charged $0.055, the same as the quote.

What went wrong: an earlier call that sent 16 generations at once timed out in the client. The tool's own guidance is never to retry a batch blindly, so we read flora_list_generations history, saw none of the 16 had started, and resent them in batches of six. A fast second poll also returned a rate-limit cooldown; waiting the 15 seconds it asked for fixed it.

Server

https://agents.flora.ai/mcp

Tools called

flora_list_models, flora_create_generations (quote, then run), flora_list_generations

Cost

$0.055, equal to the quote

Time

29.9 s

What should an MCP server for images and video do?

Look for four things beyond the generation itself.

• Price a run before it starts, in dollars.
• Run many generations in one call and read them back together.
• Pass earlier results into the next model as references.
• Save results somewhere a designer can open and keep editing.

How do I add the FLORA MCP?

In Claude on the web or Claude Desktop, add a custom connector with the URL https://agents.flora.ai/mcp and sign in to FLORA. Claude Code, Cursor, VS Code and Codex each have a one-line setup in the FLORA MCP docs at developer.flora.ai. Setup takes about two minutes.

Chat agent or coding agent?

Chat agents such as Claude and ChatGPT are easiest for one-off images and reviewing results in the conversation. Coding agents such as Claude Code, Cursor and Codex can upload local files and loop over a folder or spreadsheet, which suits catalogs and batches.

Which MCP server should your agent use for images and video?

Pick by where the work goes after the agent makes it. If a designer or a team needs to keep working on the output, a server that saves to a canvas matters most.

• Images and video for a team, priced first and saved to a shared canvas: FLORA MCP.
• Raw access to open models from a script: Replicate MCP.
• Developer pipelines with a pricing check: fal MCP.
• High-volume social and ad clips: Higgsfield MCP.
• A team already paying for Runway: Runway MCP.

Frequently asked questions