Vidofy

@vidofy/mcp

[npm] [node] [licence] [MCP]

MCP server for Vidofy — generate images, video, audio and speech from Claude, ChatGPT, VS Code, Gemini CLI, Cursor, Hermes, or any MCP client, billed to your own Vidofy account, at the same prices the website charges.

Over 570 models, including Veo 3.1, Kling 3.0, Flux 2, Seedance 2.5, Wan 2.7, Hailuo 2.3, Runway, Luma Ray 2, Qwen Image 3.0, Vidu Q3 and LTX 2 — text-to-video, image-to-video, text-to-image, image editing, video and photo effects, lipsync, text-to-speech and voice cloning. The agent browses the catalogue, prices a generation before running it, and follows one to its result.

Status: first public release.

This server is for personal Vidofy accounts. It takes one credential, VIDOFY_TOKEN, and spends your own coins — the same balance the website spends, at the same prices. There is no other billing mode: VIDOFY_API_KEY is refused at startup, and that is a decision, not a feature waiting on a release.

Tools

The usual order is list_modes → list_models → get_model → estimate_cost → generate → get_status → get_result.

generate is the only tool without readOnlyHint, which is what tells a client to ask the user before running it. It charges at submit, not on success, and returns immediately with an id — a generation takes from ~30 seconds to several minutes, so the agent polls get_status rather than holding the call open. Output is private by default; pass public: true only when the user asked for a permanent public link.

File inputs take a path on the machine running the server. The package reads the user's own file and streams it with the submit — it never makes a temporary copy — and checks the extension and size against that model's own limits first, so a file the server would reject never leaves the disk.

Not exposed, deliberately: checkout, auto top-up, purchases, referrals, the daily reward. Nothing in this package can buy coins or change a plan, however it is prompted.

Setup

There are two ways in. Take the first one unless your client cannot do it.

1. Remote connector — nothing to install

Give your client this URL:

https://vidofy.ai/mcp-app

You sign in in your browser and approve once. No token to copy, nothing to keep in a config file, and nothing to update when this package changes.

Then ask it: "list Vidofy modes" to confirm the connection, and "make me a 5-second clip of a red bicycle" — it prices the generation before running it.

Some clients cannot take this route, and it is worth knowing why before you try. Signing in here needs a client that identifies itself with a published metadata document — an https URL the authorization server fetches. Clients that instead expect to register themselves at a registration_endpoint, or to be handed a client_id and client_secret you created by hand, have nothing to work with: this server issues neither. Gemini CLI and Hermes are both in that group today. Take route 2 — it is not a lesser path, just a different way of proving who you are.

2. Local stdio server — works with every client here

Create a personal MCP token at vidofy.ai → Studio → Account → MCP Access. It is shown once.

// claude_desktop_config.json   (Cursor: .cursor/mcp.json — same shape){  "mcpServers": {    "vidofy": {      "command": "npx",      "args": ["-y", "@vidofy/mcp"],      "env": { "VIDOFY_TOKEN": "vmt_..." }    }  }}

npx fetches it on first run. Prefer a pinned copy? npm i -g @vidofy/mcp, then:

{ "command": "vidofy-mcp", "env": { "VIDOFY_TOKEN": "vmt_..." } }

The entry is the same everywhere — the file it goes in, and what the outer key is called, are not. Check yours against your client's own documentation before you paste:

So VS Code wants:

// .vscode/mcp.json{  "servers": {    "vidofy": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@vidofy/mcp"],      "env": { "VIDOFY_TOKEN": "vmt_..." }    }  }}

and Hermes wants the same thing in YAML:

mcp_servers:  vidofy:    command: npx    args: ["-y", "@vidofy/mcp"]    env:      VIDOFY_TOKEN: "vmt_..."

Both paths reach the same account, the same models and the same balance. The difference is only where the process runs and how you prove who you are.

One client shows more than the others. A generation normally comes back as text the model reads out. A host that supports MCP Apps gets a live card instead — the picture or clip itself, its progress while it runs, and a download button — and the server offers it only to a host that says it can render one. Claude and VS Code both do (in VS Code, turn on chat.mcp.apps.enabled). Everywhere else the same result arrives as text and, for an image, an inline picture. Nothing is missing; it is just quieter.

Environment

VIDOFY_API_KEY is recognised only in order to be refused: a vky_… key bills a different balance, which this server does not serve. Setting it stops startup with a message naming the token to use instead — and setting both is refused too, since the two bill different balances and no precedence rule is worth having to remember.

Development

npm installnpm run buildnpm run inspect      # MCP Inspector — spends nothing

VIDOFY_API_BASE points it at a different origin, if you are running one.

Nothing here writes to stdout. With stdio transport, stdout is the protocol channel — a single stray console.log() puts a non-JSON line in the stream and the client drops the connection with an error that explains nothing. Diagnostics go to stderr via the log() helper in src/index.ts.

Layout

src/config.ts          credential + mode + base URL, validated at startupsrc/backend.ts         the only place that talks HTTP: auth, retries, multipart, errorssrc/schema.ts          one model's m_options → a JSON Schema the agent can fill insrc/map/b2c.ts         both response shapes → one; strips the provider costsrc/tools/info.ts      list_modes, list_models, get_modelsrc/tools/generation.ts estimate_cost, generate, get_status, get_resultsrc/tools/account.ts   get_balance, get_usagesrc/index.ts           the server: stdio transport, tool registration, annotations

server.json     MCP registry manifest (name must match package.json "mcpName")

Licence

MIT