fal.ai MCP Server: Image & Video Generation Guide
How to set up fal.ai's MCP server for image and video generation, what its 11 tools return, real per-run costs, and where the pipeline stops.
fal.ai runs an official MCP server that puts more than 1,000 image and video models inside your coding agent. It is free to connect, you pay per model run, and your agent can check a model’s price before spending anything. The two things to know before you wire it up: it does not yet work in Claude on the web or Claude Desktop, and what comes back is a raw clip or image, not a finished video.
Key takeaways
| Question | Short answer |
|---|---|
| Is there an official server? | Yes, hosted by fal at https://mcp.fal.ai/mcp |
| Do I need an API key? | Yes. It goes in an Authorization header |
| Does it do image and video? | Yes, both, from one server |
| Where does it work? | Claude Code, Cursor, Windsurf. Not Claude web or Desktop yet |
| What does it not do? | Script, voiceover, captions, editing, publishing |
What the fal.ai MCP server is
fal.ai is a generative media platform for developers: image, video, and audio models running on serverless GPUs, billed per use. Its official MCP server, hosted at https://mcp.fal.ai/mcp, exposes that whole catalogue to any MCP client over streamable HTTP.
The server ships 11 tools in three groups:
- Discovery:
search_models,get_model_schema,get_pricing,search_docs, andrecommend_model. Your agent can find a model, read its inputs, and check its price before running anything. - Execution:
run_modelfor fast jobs like images, plussubmit_job,check_job,get_job_result, andcancel_jobfor long video generations. - Utility:
upload_file, for feeding your own image or audio into a model.
The discovery tools are the honest selling point. Most video MCP servers make your agent generate blind. Here it can compare three models and their prices first, then spend. MCP itself is an open standard; we cover the wider field in our roundup of the best MCP servers for AI video generation.
How to set it up
You need a fal account and an API key from the fal dashboard. The key rides along in an Authorization header on every request; fal stores nothing server-side.
Claude Code is one command:
claude mcp add --transport http fal-ai https://mcp.fal.ai/mcp --header "Authorization: Bearer YOUR_FAL_KEY"Cursor takes the same URL and header in mcp.json, then a restart. Windsurf takes them under Settings, then MCP.
That header is also the main thing to handle with care. Your fal key can spend real money, so treat the config file like a password. This is a real difference from sign-in based servers, where revoking access is one click in your account.
One server, images and video together
fal’s server is unusual in covering both halves of creative generation. In one session your agent can:
- Generate a set of stills with an image model like Seedream V4 or Flux Kontext Pro, using
run_model. - Feed one of those stills into a video model with
upload_file, for an image-to-video shot. - Kick off a Veo 3, Kling, or Wan clip with
submit_job, keep working, and collect it later withget_job_result.
The job tools matter for video. Clips take minutes to render, and a stateless submit-and-poll flow means your agent is not stuck waiting on a timeout. It is a genuinely well-designed developer surface.
What you get back, and what you do not
Here is where expectations need setting. fal is fast model inference: prompt in, raw asset out. A publishable video takes six steps, and fal’s server covers one of them.
You get: a silent clip or an image, from whichever of the 1,000+ models your agent picks, delivered as an asset URL. Plus the schema and price data to pick well.
You do not get: a written script, a voiceover, captions, an edit that joins clips, or a publish step to YouTube or TikTok. fal does not pretend otherwise; it is a developer inference platform, not a video studio.
For a developer testing which model fits a product, that trade is exactly right. For a creator who wants to ask Claude for a video and post the answer, it is five steps short.
What it actually costs
The server is free. You pay fal’s normal per-run rates for each model you trigger, and there is no subscription. These were the listed prices for popular models on fal’s pricing page when we checked on 31 August 2026:
| Model | Type | Price |
|---|---|---|
| Wan 2.5 | Video | $0.05 per second of output |
| Kling 2.5 Turbo Pro | Video | $0.07 per second |
| Veo 3 | Video | $0.40 per second |
| Seedream V4 | Image | $0.03 per image |
| Flux Kontext Pro | Image | $0.04 per image |
So an 8 second Wan clip runs about $0.40, the same clip on Kling about $0.56, and on Veo 3 about $3.20. An idle month costs nothing, which makes this one of the cheapest ways to experiment across models.
fal.ai MCP vs AITuber MCP
Both are official, hosted servers. They solve different problems.
| fal.ai MCP | AITuber MCP | |
|---|---|---|
| Returns | Silent clip or image, as an asset URL | Finished MP4 with voice and captions |
| Model choice | 1,000+ models, agent can compare | Handled for you |
| Script writing | No | Yes |
| Voiceover | No | Yes, 1,300+ voices |
| Captions | No | Yes, word synced |
| Publishing | No | YouTube, TikTok, Instagram, Facebook, Threads, X |
| Auth | API key in a header | Account sign-in |
| Claude web and Desktop | Not yet | Yes |
| Pricing | Pay per run, no plan | Free credits, then plans from $29 a month |
The honest read: fal wins on model breadth, price transparency, and pay-as-you-go economics. AITuber wins when the thing you want back is a video you can post without opening an editor. They also stack; nothing stops one agent from using fal for a hero shot and a pipeline server for the finished piece.
Who should use fal.ai MCP
Use it if you write code, you want to test several image or video models before committing, or you need image and video generation in one agent workflow with no subscription.
Skip it if you work in Claude on the web or Claude Desktop, since the server does not connect there yet. Or if what you need each time is a publishable video with narration and captions, because fal returns the raw material, not the product.
Use both if you prototype shots in fal from Claude Code and let a pipeline server assemble the videos your channels actually post.
Frequently Asked Questions
Basics
Is there an official fal.ai MCP server?
Yes. fal hosts it at https://mcp.fal.ai/mcp and documents it on fal.ai. It uses streamable HTTP transport with your fal API key in an Authorization header on each request.
Does the fal.ai MCP server do both images and video?
Yes, and that combination is its standout feature. Image jobs run through run_model; video jobs use submit_job with check_job and get_job_result polling, so long renders do not time out.
How many models does it expose?
fal’s catalogue is over 1,000 models, including Veo, Kling, Wan, Seedream, and Flux. The search_models and recommend_model tools help your agent choose.
Setup
How do I add fal.ai MCP to Claude Code?
One command: claude mcp add --transport http fal-ai https://mcp.fal.ai/mcp --header "Authorization: Bearer YOUR_FAL_KEY". Get the key from your fal dashboard first.
Does fal.ai MCP work in Claude on the web or Claude Desktop?
Not yet. Those clients require OAuth sign-in for remote connectors, and fal’s server authenticates with an API key header. fal’s docs list Claude Code, Cursor, and Windsurf as the supported clients.
Is my API key stored on fal’s server?
No. The key is passed per request in the Authorization header, and fal states nothing is stored server-side. The key does live in your local client config, so protect that file.
Cost
How much does fal.ai MCP cost?
The server is free. You pay standard fal per-run pricing for each model call, for example $0.05 per second of Wan 2.5 video or $0.03 per Seedream V4 image at the time of checking. There is no subscription and idle months cost nothing.
Can my agent check a price before generating?
Yes. The get_pricing tool returns a model’s rate, and get_model_schema returns its inputs. That pre-flight check is the server’s best guard against surprise bills.
Comparisons
Does fal.ai MCP add voiceover and captions?
No. It returns silent clips and images. Script, narration, captions, editing, and publishing stay with you. A full pipeline server like AITuber returns those steps finished.
fal.ai MCP or the fal API directly?
Same platform, same pricing. Use MCP when an agent is driving the work in conversation. Use the API when your own code is. Our MCP vs API guide covers the trade in depth.
How we researched this: server URL, transport, tool names, client support, and the OAuth limitation come from fal.ai’s own MCP documentation, and per-model prices come from fal.ai’s pricing page, both read on 31 August 2026. MCP ecosystem figures come from Anthropic’s published announcement. AITuber is a product we build, so we have flagged every comparison and kept the fal details accurate and current. fal’s catalogue and prices change often; check fal.ai for the latest.