GitHub
AI Media RouterFree · Open source

One MCP server for AI image and video generation.

Your agent asks once. Seedance, Kling, Veo, FLUX or GPT Image, at whichever of 43 providers you pick, with your own keys.

Get it on GitHubWhy prices matter
  • macOS · Windows · Linux
  • MCP · REST · CLI
  • No account, no markup
  • MIT on GitHub

Source: github.com/jutsuhq/ai-media-router

ai-media-router mcp
Your AI agent (Claude Code, Cursor, Codex) asks AI Media Router once; the router checks price and budget, fills in references and submits to one of 43 providers with your key; the result is saved to your local library and the agent gets back a local file path.
Runs on your machine · your keys43

The same clip, half the price, at another provider.

AI Media Router is an open-source MCP server for AI image and video generation. Before your agent spends anything, search_prices shows every offer for the model, cheapest first, so it can pick. For Seedance 2.5 at 720p with audio, a 10-second clip ranged from $2.16 to $4.62 among the providers the router can call on November 1, 2026.

ProviderPer second10 s clip
302.AI$0.216$2.16
Volcengine (ByteDance, China) · official$0.226$2.26
Runware$0.230$2.30
BytePlus (ByteDance) · official$0.231$2.31
Replicate$0.231$2.31
kie.ai$0.315$3.15
WaveSpeedAI$0.324$3.24
fal$0.462$4.62

Seedance 2.5 text-to-video, 720p, with audio: list prices from ai-media-router prices "seedance 2.5" --res 720p, run November 1, 2026; eight of the 23 providers the router serves for this model. The catalog is jutsuhq/ai-media-api-prices, the same data as our AI video API prices and AI image API prices, updated daily. Today's numbers for this model: Seedance 2.5 API prices.

From one request to a file on disk.
The agent decides; the router checks.

  1. 1

    Compare

    search_prices lists video offers per second and image offers per image, with the provider key to call.

  2. 2

    Check

    preflight runs every check, keys, inputs, cost and budget, and sends nothing.

  3. 3

    Generate

    generate submits with your key. A worker follows the job, even across restarts.

  4. 4

    Keep

    The result is downloaded, hashed, previewed and indexed before the provider's link expires.

  5. 5

    Reuse

    Named references keep a character or product the same; one job's last frame starts the next shot.

  6. 6

    Finish

    Trim, concat, reframe, overlay and add audio with ffmpeg on your machine, at no cost.

43 providers, one set of tools.

Each provider is a TOML file with recorded fixtures. When an API changes, an agent can read the raw exchange, fix the file and save it, with no new release.

KindProviders
Inference platforms and API aggregatorsfal, Replicate, Runware, WaveSpeedAI, kie.ai, OpenRouter, Vercel AI Gateway, Together AI, DeepInfra, Novita AI, SiliconFlow, Segmind, Atlas Cloud, each::labs, AI/ML API, 302.AI, APIMart, CometAPI, EvoLink, PiAPI, ModelsLab, Makernice, Pollo AI, Sume, Unifically, Yir, Higgsfield, Comfy partner models
Model makers' own APIsBytePlus and Volcengine (Seedance, Seedream), Kling, Gemini API and Google Vertex AI (Veo, Imagen, Nano Banana), Alibaba Cloud (Wan), MiniMax (Hailuo), Runway, Luma, LTX, PixVerse, Vidu, xAI (Grok Imagine)
Self-hostedYour own ComfyUI server or a rented GPU: any workflow in API format, 0 USD per job

Models are not a fixed list: the model name goes to the provider as written. Specs not yet checked against a live response say so to the agent; live checks of every spec are on the project's roadmap.

Compared with fal MCP, Replicate MCP and ComfyUI.

FeatureAI Media RouterOne provider's MCP serverA local ComfyUI
Providers43, plus your ComfyUIThat oneModels you run yourself
PricesCompared across providers before the agent picksIts ownFree per job, your GPU
Spending limitUSD budgets, checked before a job is sentWhat that provider offersNot needed
ResultsA local library: search, ratings, references, lineageThat provider's storage or linksComfyUI's output folder
AccountNone: your keys go straight to each providerThat provider'sNone

fal and Replicate each run an MCP server for their own platform (fal, Replicate), and both are among the router's providers. If you only ever use one provider, its own server is enough. Not affiliated with fal, Replicate, OpenRouter or Comfy Org.

Add it to Claude Code in one line.

Download the binary for your system from the releases page, put your provider keys in the environment or in config.toml, then connect it. Claude Desktop, Cursor and Codex take the same ai-media-router mcp command.

claude mcp add ai-media-router -- ai-media-router mcp

Budgets

A named USD cap refuses a job before it is sent, and a batch fits whole or not at all.

No paying twice

An identical request already running is joined, not sent again, when an agent retries after a timeout.

Provenance in the file

Prompt, provider, model and parents are written into each PNG and JPEG, and into MP4 with ffmpeg. Keys are redacted.

Prices and tools behind the router.

What video and image APIs cost per provider, the Higgsfield ways in for agents, and running models on your own machine.

Ten seconds of video with sound at 1080p: Kling 4.0 $1.65, Veo 3.1 $4.00, Seedance 2.5 $5.69
SeedanceSeedance 2.5 vs Kling 4.0 vs Veo 3.1: Price and LimitsSeedance 2.5, Kling 4.0 and Veo 3.1 side by side: API price per second at 720p, 1080p and 4K, clip length, references and video editing, checked October 31.
Three cards comparing the Higgsfield API, billed from a dollar balance, with MCP and the CLI, which spend plan credits
HiggsfieldHiggsfield API vs MCP vs CLI: Which to UseWhat Higgsfield's API, MCP server and CLI each are, how to connect them to Claude or your code, and why the API bills dollars while MCP and CLI spend credits.
A Higgsfield API job from reserve to submit, webhook or poll, store and settle, next to the note that the prepaid balance is the only hard spending cap
HiggsfieldHiggsfield API in Production: Webhooks, Limits, Spend CapsHow the Higgsfield API behaves in production: webhook signatures, idempotency keys, concurrency limits and 429s, and the spending cap it does not have.
1 hour 22 minutes to render 2 seconds of 480p Wan 2.2 14B in ComfyUI on an M1 Max with 64 GB
MacComfyUI on Mac (Apple Silicon): Install, Requirements, SpeedInstall ComfyUI on an Apple Silicon Mac and generate your first image. Desktop setup, model downloads, memory requirements, MPS limits and video options.
A ComfyUI graph: Load Image and Load Video nodes feeding the Seedance 2.5 partner node and Save Video
SeedanceSeedance 2.5 in ComfyUI: Setup, Credits and CostHow the Seedance 2.5 partner node works in ComfyUI: what it needs, how Comfy credits are billed, what a clip costs, and local ComfyUI vs Comfy Cloud.
Unified memory open video models ask for on a Mac: HunyuanVideo 1.5 in MLX at Q4 about 14 GB, LTX Desktop 32 GB for 540p and 48 GB for full quality, SkyReels V2 14B 51 GB at 540p
MacLocal AI Video Generator on a Mac: Models, Memory, AppsWhich open AI video models run locally on Apple Silicon: LTX-2, HunyuanVideo, Wan 2.2, FramePack and SkyReels. Memory per Mac, published speeds, apps.

FAQ.

What is AI Media Router?

A free, open-source (MIT) MCP server that lets Claude, Cursor, Codex or any MCP client generate images and videos through 43 providers, such as fal, Replicate, kie.ai, Runware, OpenRouter, BytePlus and Kling, or your own ComfyUI server. It runs on your machine as one Rust binary, uses your own API keys and saves every result to a local library.

Is there an MCP server for AI image and video generation across providers?

Yes, AI Media Router is one. Providers such as fal and Replicate run their own MCP servers for their own platforms; AI Media Router puts 43 providers behind one set of tools (search_prices, generate, get_job) and adds price comparison, budgets and a local library.

Does it pick the cheapest provider automatically?

No. search_prices lists offers cheapest first from a price catalog updated daily, and the agent or you choose. There is no silent routing or fallback. On November 1, 2026 the catalog listed Seedance 2.5 at 720p with audio from $0.216 a second (302.AI) to $0.462 (fal) among the providers the router can call.

Is it free?

Yes. You pay each provider directly at its own prices with your own key. The router has no account, no balance and no markup, and ComfyUI jobs on your own server cost nothing.

Is it like OpenRouter for images and video?

In the sense of one interface over many providers. Unlike OpenRouter it runs on your machine and has no account or balance: each provider bills you directly. OpenRouter is one of the 43 providers it can call.

Which models can it use?

Any model a supported provider serves, for example Seedance, Kling, Veo, Wan, Hailuo, Runway, Luma, LTX, FLUX, GPT Image, Nano Banana, Imagen and Seedream. The model name and input go to the provider unchanged, so a model a provider adds today works today.

Can an agent overspend?

Not with a budget set. A named USD budget refuses a job before anything is sent when what was spent plus the job’s estimate would exceed it. The check works from estimates, so a provider that bills more than its listed price can still go over by that difference.

Does it work with ComfyUI?

Yes. Set COMFYUI_URL and your server becomes the provider comfyui: the input is an API-format workflow, and library files used in a LoadImage node are uploaded for you. Comfy’s hosted partner models are a separate provider.

Which systems does it run on?

macOS (Apple silicon and Intel), Windows x86_64 and Linux x86_64 and arm64, as a prebuilt binary, a container image or built from source with Rust 1.89 or later. It speaks MCP over stdio or HTTP, plus a REST API and a CLI.

AI Media Router logo

Give your agent every image and video API.

Free and open source. One binary, your keys, your files.

Get it on GitHubQuick startVideo API pricesImage API prices

macOS, Windows and Linux. Independent project. Not affiliated with any provider it calls.