Local files → URLs

No MCP tool accepts a filesystem path

Every file input across the whole MCP contract — imageUrls, videoUrl, the media tools' asset references, picsart_drive's url — is an HTTP(S) URL. There is no filePath parameter anywhere. /Users/me/photo.jpg will never work.

Some gen-ai tools additionally accept an inline data: URI. Both the gen-ai MCP server and Picsart Media Studio also have a built-in uploader that produces the URL for you, which is route D below.

This is not an oversight. An MCP server runs somewhere else (Picsart's infrastructure); it has no access to the filesystem of the machine the agent is running on. So before any local file can be used as a generation input, something on your side has to give it a URL.

Pick by which connector you are using, and by what your agent host can do.

The easiest route: the built-in uploader

Ask your agent to open the Picsart uploader (route D below). It works on both the gen-ai MCP server and Media Studio, and needs no shell or CLI.

Comparison

Path	Needs	Good for	Cost
A. Picsart CLI upload	Shell access	Anything — the general answer	Free, no tokens
B. Chat attachment	A host that forwards attachments (e.g. the ChatGPT app)	Files the user drags into the chat	Free, no tokens
C. data: URI	Nothing	Small images only, as a stopgap	Very expensive in tokens
D. Built-in uploader (picsart_media_upload)	The gen-ai MCP server or the Media Studio connector, in a host that shows interactive panels	Any local file, dropped in from your browser	Free, no tokens

A. Shell-capable agent → the Picsart CLI

If your agent can run shell commands (Claude Code, Cursor, Codex, a CI job) and you have the CLI installed, you can upload with it, then hand the resulting URL to any MCP tool. The CLI is optional: MCP itself never needs it.

gen-ai upload ./photo.jpg                 # single file
gen-ai upload ./renders/ -r               # a folder, recursively
gen-ai upload ./photo.jpg -f "Campaign"   # into a named Drive folder

Add --json to get the resulting URL back — gen-ai upload prints { ok, files: [{ path, url, driveUid, error }] }:

gen-ai upload ./photo.jpg --json | jq -r '.files[0].url'

For a single file, gen-ai upload-to-drive always prints a one-line JSON result:

gen-ai upload-to-drive ./clip.mp4
# {"status":"ok","drive_url":"https://cdn.../clip.mp4","drive_uid":"...","file_name":"clip.mp4","elapsed_ms":812}

It sets the Drive resource type (photo, video, or audio) from the file extension; --name sets the display name and --folder the Drive folder.

Then pass the URL straight through:

{ "name": "picsart_remove_bg",
  "arguments": { "image": "https://cdn.picsart.com/.../photo.jpg" } }

The CLI requires gen-ai login once, see Authentication. This sign-in is for the CLI only.

B. Chat attachment → picsart_drive upload

Hosts that expose user attachments to MCP tools (notably the ChatGPT app) can pass the attachment handle directly to picsart_drive. Nothing has to touch a filesystem:

{ "name": "picsart_drive",
  "arguments": { "action": "upload", "file": "<attachment>" } }

The result's result.url is a CDN-hosted URL, ready to pass to picsart_generate's imageUrls or any media tool. Add folderUid to choose a destination folder and type to set the resource kind.

This is the best experience where it's available — the user drags a file into the conversation and the agent does the rest. It is not available in hosts that keep attachments out of tool arguments, which includes most desktop MCP clients today.

C. data: URI → picsart_drive upload

picsart_drive's upload action accepts an inline data: URI in its url parameter (the server pushes it to the Picsart CDN first and returns the resulting URL). So an agent that can read the file into its own context — but has no shell — can inline it:

{ "name": "picsart_drive",
  "arguments": {
    "action": "upload",
    "name": "logo.png",
    "url": "data:image/png;base64,iVBORw0KGgoAAAANS..."
  } }

This costs real model-context tokens

Base64 inflates the file by ~33%, and every character of it passes through the model's context window. A 1 MB file is roughly 350,000 tokens — likely more than the context window allows, and billed as model input on the agent's side even when it fits.

Use this only for small images (icons, logos, thumbnails — tens of KB), and only when neither path A nor path B is available. It is a stopgap, not a general solution. Never do this for video.

What about remote URLs behind auth?

If the asset already lives on a URL the render service can't reach (a short-lived signed URL, or a host requiring auth), copy it into Drive first with the same upload action, passing the remote URL instead of a data: URI:

{ "name": "picsart_drive",
  "arguments": { "action": "upload", "name": "ref.jpg", "url": "https://example.com/ref.jpg" } }

The returned CDN URL is stable and publicly fetchable by the generation and render services.

D. The built-in uploader

The gen-ai MCP server and Picsart Media Studio both include an uploader (picsart_media_upload). Ask your agent to open it, drop the file in, and carry on. It uploads straight from your browser, and you can also pick a file already in your Picsart Drive.

Nothing to install, and no URL to produce yourself. It needs a host that can show interactive panels in the conversation, such as Claude or ChatGPT.

More

• Files & Drive — the rest of the Drive surface
• MCP Quickstart — the full tool catalog
• CLI Quickstart — installing and using gen-ai