fal
Generate images, videos, audio and 3D with fal. Discover models, inspect schemas and pricing, and run jobs through a hosted MCP connection with browser OAuth.
Hosted MCP Server
npx add-mcp 'https://mcp.fal.ai/mcp-relay'Installs into Claude Code, Codex, Cursor and more
Documentation
fal MCP Server Connect your AI to 1,000+ models
A hosted endpoint that lets any AI assistant search, run, and chain more than 1,000 generative AI models directly from a conversation — image, video, audio, 3D, upscaling, and more. One command. No SDK. No docs to read.
Setup
Up and running in two steps
The fal MCP Server is hosted — nothing runs locally. One command adds it to your coding agents, and apps connect in a few clicks. Sign in with your fal account and start generating.
1. Add the fal MCP server
Coding agents
One command sets up every coding agent on your machine
curl -fsSL https://fal.ai/mcp/install.sh | bash
- Claude Code
- Codex CLI
- Cursor
- VS Code
- Devin Desktop (Windsurf)
- Gemini CLI
- Grok Build
- opencode
Run it in a terminal on macOS or Linux. It finds which of these agents you have, adds fal to each, and prints the one sign-in step each agent needs. Guides for every agent are in the docs.
Or set up one agent
Claude Code
claude mcp add --transport http --scope user fal https://mcp.fal.ai/mcp-relay
Then run /mcp, select fal and choose Authenticate.
Codex CLI
codex mcp add fal --url https://mcp.fal.ai/mcp-relay
Codex opens the browser sign-in itself.
Cursor
Add the plugin, or run /add-plugin fal-ai in Cursor. It adds the fal MCP server and fal's media skills. Then open Customize, find fal and sign in.
Only want the MCP server? Add it without the plugin.
Setting up from inside a coding agent?
Paste this into the agent. It uses the agent's own MCP command, which agents can run without approving an install script.
Add the fal MCP server to this coding agent with its own MCP command, not an install script. The server is https://mcp.fal.ai/mcp-relay (remote, Streamable HTTP, OAuth sign-in). In Claude Code: claude mcp add --transport http --scope user fal https://mcp.fal.ai/mcp-relay. In Codex: codex mcp add fal --url https://mcp.fal.ai/mcp-relay. Then tell me how to sign in.
Apps and chat agents
Connect fal inside the app
- Open the fal plugin page. Click Install plugin.
- Select Connect for the fal app if prompted. Complete browser sign-in and authorization.
- In a new chat, type
@fal. Install fal in ChatGPT →
2. Ask for anything
Start a new chat or session once you have signed in, and ask:
“Generate a cinematic video of a fox running through snow, then upscale it to 4K.”
That's it. Your assistant now has access to every model on fal — it searches the catalog, checks pricing, picks the right model, and runs it, all in one turn.
Documentation
Read the MCP docs
Setup guides for every client, the full tool reference, and troubleshooting tips.
Dashboard
MCP & connectors
Choose which fal account your assistant bills: your personal account or a team.
Community
Join the fal Discord
Share what you build, get help from the team, and see what others are running through MCP.
Tools
Tools, zero configuration
The MCP server exposes tools that your assistant picks automatically. You never call them directly — you just describe what you want.
Discover
search_models
Search the catalog by keyword or category.
recommend_model
Describe the task; get candidate models and published guidance.
get_model_schema
Read a model's exact inputs and outputs before running it.
search_guidance, read_guidance, search_docs
Read fal's prompting and model-routing guidance and docs.
Price and account
get_pricing
Unit prices for up to 50 models in one call.
estimate_cost
Estimate one request from its real inputs. Advisory, not a quote.
list_model_sets
Use your Sandbox model sets for side-by-side comparisons.
check_account_status, list_accounts, select_account
See which fal account is billed and switch to a team you belong to.
Generate
run_model
Run a model and wait a bounded time for the result.
submit_job, submit_jobs
Queue long video and 3D jobs, or up to 8 requests at once.
check_job, wait_for_jobs
Follow queued jobs without resubmitting them.
get_job_result, cancel_job
Fetch finished output, or cancel a job you no longer need.
Your inputs
upload_file
Upload from a public URL, base64, or a signed upload up to 90 MB.
upload_attachment, verify_upload
Import a chat attachment where the app allows it, and confirm the upload.
inspect_images
Look at up to 4 existing images before editing them.
Workflows
prepare_media_workflow
Check up to 8 chained steps against live schemas before spending credits.
prepare_media_transform
Plan an upscale, cutout, HDR delivery or video outpaint.
resolve_model_inputs
Turn saved assets, characters and past jobs into valid model inputs.
prepare_deliverables
Check up to 32 outputs against the delivery you asked for.
Library and history
save_job_outputs, save_asset
Save results to your fal library, keeping where they came from.
search_assets, list_collections
Find saved media and organize it into collections.
create_entity, list_entities
Keep reusable characters, props, environments, styles and scenes.
search_request_history, get_job_recipe
Find past generations from the last 90 days and their exact settings.
show_media, open_library
Preview results and browse your library in apps that display MCP Apps, such as ChatGPT.
Use cases
What can you do with it?
Real prompts, straight from a conversation. Copy one into your assistant and watch it pick the models, run them, and hand back the results.
Image generation
Generate images with the best model
“Generate a photorealistic portrait of a woman in a wheat field at golden hour”
Your assistant searches the catalog, finds the most suitable model, checks the parameters, and runs it — all in one turn. The assistant picks the model. You describe the result.
LLM chaining
Chain LLM + image generation
“Expand this idea into a detailed prompt and generate the image: a cozy Japanese coffee shop in the rain”
An LLM writes a detailed prompt with lighting, camera angle, and mood — then the image model generates with it. Dramatically better than a simple prompt.
Image to video
Turn a photo into a video
“Take this photo and turn it into a 5-second cinematic video with a slow zoom”
Upload → model selection → queue management → result. All handled automatically.
Creative pipeline
Run a full creative pipeline
“Create a product ad for Aurora headphones: write a tagline, generate a product image, and create a voiceover”
Three models chained in one conversation: an LLM writes the copy, an image model generates the visual, a TTS model reads the tagline. You get all the URLs back.
Model comparison
Compare models side by side
“Compare two image models for a minimalist logo on white background”
Run both, see both results. Model benchmarking in one sentence.
Pricing
Check pricing before you run
“How much does it cost to generate images with FLUX and videos with Kling 3.0?”
Know the cost before you spend.
Multi-step workflows
Chain complex workflows
“Generate a futuristic Tokyo street, upscale it to 4K, and describe what's in it using a vision model”
Chain generation → upscaling → analysis in a single conversation.
How it works
Stateless, isolated, and free to use
Your assistant sends a request to mcp.fal.ai signed in with your fal account, the server calls the fal Platform API on your behalf, and the results come straight back to your conversation.
Hosted and stateless
The server is fully stateless and hosted on Vercel. No sessions, no state — each request is completely isolated.
No API key to paste
Your client signs in to your fal account in the browser. There is no key to create, paste into a config file, or rotate.
Free to use
The MCP server is free. You only pay for the model runs you trigger, at standard fal pricing.
FAQ
Common questions about the fal MCP Server
What is the fal MCP Server?
The fal MCP Server is a hosted endpoint at https://mcp.fal.ai/mcp-relay that lets any AI assistant search, run, and chain 1,000+ generative AI models on fal directly from a conversation. It is built on the Model Context Protocol (MCP), an open standard that lets AI assistants use external tools. Your assistant doesn't just generate code that calls fal — it actually calls fal.
Which AI assistants are supported?
Any MCP-compatible client with OAuth sign-in works. ChatGPT, Claude, Claude Code, Codex, Cursor, Grok Bot, Muse, Instinct and Replit have setup steps on this page; Cursor and Grok Bot install fal as a plugin from the Cursor Marketplace. Other clients connect to the same server URL.
How do I add fal to all my coding agents at once?
Run curl -fsSL https://fal.ai/mcp/install.sh | bash on macOS or Linux. It adds the fal server to Claude Code, Codex CLI, Cursor, VS Code, Devin Desktop (Windsurf), Gemini CLI, Grok Build and opencode, skips any that are already set up, and writes no API key. Then complete browser sign-in once in each agent.
Do I need an API key?
No. Your client signs in to your fal account in the browser, so there is no key to create or paste. Only clients that can't sign in with MCP OAuth need a key: send it as Authorization: Key <key_id>:<key_secret> to the same URL.
I set it up with an API key before. What changed?
The old setup used an API key in an Authorization header and now fails with 401. Remove that fal entry from your client and add https://mcp.fal.ai/mcp-relay again without a header, then sign in in the browser.
How much does it cost?
The MCP server itself is free. You only pay for the model runs your assistant triggers, at standard fal pricing — the same rates as calling the API directly. Your assistant can even check the price of a model with the get_pricing tool before running it.
Which fal account gets billed?
Your Active MCP account: your personal account by default. To use a team's credits, select the team in MCP & connectors (fal.ai/dashboard/ai-tools). New requests switch without reconnecting.
What tools does the server expose?
Tools for finding models (search_models, recommend_model), checking prices (get_pricing, estimate_cost), running and queueing generations (run_model, submit_job, check_job), uploading inputs (upload_file), and saving results to your fal library. Your assistant picks them automatically — you just describe what you want.
What kinds of models can I run?
The entire fal catalog: image generation, video generation, audio and text-to-speech, 3D, upscaling, LLMs, vision models, and more — over 1,000 models in total.
Can my assistant chain multiple models together?
Yes — that's one of the biggest advantages. A single request like "generate a futuristic Tokyo street, upscale it to 4K, and describe what's in it" chains generation, upscaling, and analysis in one conversation. The assistant handles uploads, queueing, and passing results between models.
Can I use my own images?
Yes. Give your assistant a public URL, or attach the file in an app that shares attachments with tools. The hosted server can't read files on your computer.
The fal tools don't show up. What should I check?
Finish the browser sign-in, then start a new chat or session. In Claude, enable fal from the + menu under Connectors; in ChatGPT, type @fal. Check that the server URL is https://mcp.fal.ai/mcp-relay and that no Authorization header is set.