Skip to main content
Modellix Media Generation provides a remote Streamable HTTP MCP endpoint for asynchronous image, video, and audio generation. It uses MCP 2026-07-28 with per-request metadata and requires a client that supports this protocol revision.

​
Connect a Compatible Client

Configure a remote HTTP MCP server with the following URL and an Authorization: Bearer <your Modellix API key> header:
For clients whose HTTP MCP configuration uses mcpServers, url, and headers, the equivalent configuration is:
Use your client’s secret store or supported environment substitution for the key. Do not commit the filled configuration. Check the client’s protocol support before using this example; HTTP configuration fields vary by host. The server is stateless. It does not require a sticky session or an initialization handshake from an older MCP revision. Discovery uses server/discover; requests carry MCP-Protocol-Version, Mcp-Method, the applicable Mcp-Name, and protocol/client-capability metadata. The client should construct these fields. Older clients receive a protocol error instead of silently selecting another version.

​
Generate Media

  1. Call modellix_list_models with an optional type or keyword. Use limit and offset to page through results; the default limit is 20 and the maximum is 100.
  2. Read modellix_get_schema using the exact provider/model slug. This public request needs no API key and preserves required fields, defaults, enums, and schema references.
  3. Upload an input only when the selected model requires one, or reuse an existing file URL.
  4. Confirm the model and input with the user, then call modellix_submit_task once. Keep the returned task_id.
  5. Call modellix_wait_task for up to 45 seconds per call. A normal deadline returns a snapshot; continue waiting with the same ID.
  6. Read successful outputs from data.result.resources and save them before they expire.
modellix_get_schema returns the original API schema document with an input_schema projection and a reference_scope explanation. Resolve input-local references against input_schema, and document-scoped references against the original returned document. Keep the document root when its references use paths, post, or components; copying only the projection can break those references.

​
Upload and Reuse Inputs

Pass exactly one of url and base64_content. Base64 uploads also require a plain filename. Remote MCP cannot read a path on your computer. URL downloads accept public HTTP(S) destinations and reject private, local, reserved, or credential-bearing targets, including redirects. MCP uploads are limited to 16 MiB (16777216 bytes). The REST API has a separate configurable limit. Supported extensions are JPG, JPEG, PNG, WebP, GIF, BMP, TIFF, TIF, HEIC, HEIF, MP4, M4V, WebM, MOV, MKV, AVI, MP3, WAV, M4A, AAC, OGG, FLAC, OPUS, and WEBA. The media API still validates content and team policy. Default team settings allow 10 files, 2 concurrent uploads, and about 7 days of retention; your configuration may differ. On an upload quota error, list files, select an unneeded file_id, and explicitly approve its deletion. The tools never automatically delete files or retry uploads to free quota. Deleting an absent file succeeds with already_absent.

​
Recover Without Duplicate Submissions

modellix_submit_task and uploads are sent once. If the connection fails before a trustworthy response is received, submission_unknown means the operation may have succeeded. Inspect existing tasks or files before any manual retry. An unknown submission cannot guarantee a task ID. modellix_get_task returns immediately. modellix_wait_task defaults to 30 seconds and accepts 1–45 seconds; a normal timeout is HTTP success with data.wait.timed_out=true, not a failed task. Terminal states are success, failed, and canceled. The media endpoint does not advertise the official MCP Tasks extension, Apps UI, or enterprise identity-provider login. It does not choose a default model or estimate a price on your behalf. The Docs MCP searches documentation. The Plugin keeps its CLI-first execution workflow. Agent Canvas remains a local stdio product.

REST API

Submit async image, video, and speech tasks over HTTP.

Skill

Teach coding agents how to discover models and run generation.

CLI

Authenticate, submit tasks, wait, and download results from the terminal.

Playground

Generate and preview media in the browser with no code.