MCP limitations
Rate limits of the Slicer MCP server, how the assistant transfers files, waiting for results and features that are available only in the app for now.
On this page
The MCP server works on top of the same API as the REST API, so it has the same limits and the same version 1.7 capabilities. Here's what matters when working with an assistant.
Rate limits
Limits are counted per connection (for an API token — per token), as in the REST API:
| What | Limit |
|---|---|
| Requests to the MCP server, including tool calls | 120 per minute |
| Concurrent imports from URLs | 3 per account |
Source frames per get_frames call | 16, 512 pixels on the longer side |
Observations from the frames (scenes) per submit_director_plan plan | 400 |
Assistants don't reach these limits in normal work. If a limit is exceeded, the server answers 429 with a Retry-After
header — the client waits and retries. Details — Rate limits.
Files
The MCP protocol carries text and data, not large files, so a video gets into Slicer outside the conversation with the model:
- an assistant that can run commands on your computer calls
create_upload_linkand sends the file itself withcurl -T. The link is single-use, lives at most an hour and accepts one file of the declared size; - an assistant without access to your files can only upload a file from a public
httpslink withimport_from_url. The link must point directly to the file, not to a video hosting page; - files you attach to a chat message stay with the AI client — Slicer doesn't see them. Upload them in the Slicer app or give the assistant a link.
Formats, sizes and the number of materials are the same as in the app: File formats and limits.
Waiting for results
A project without the AI director becomes a draft within seconds, and edits apply immediately. Export, captions from speech
and an AI director montage (usually 3–6 minutes) take longer. get_project with waitSeconds (up to 50 seconds) waits on
the server and returns as soon as the project's status, stage or progress changes.
Speech recognition progress is shown by get_transcribe_jobs. A few such calls are enough instead of frequent polling.
Some clients limit how long a single tool call may take — then the assistant simply calls the tool again.
The finished video
The assistant sends a link to the project in the web app (appUrl): there you can watch and download the video after signing
in. For programs, the file is available at /v1/projects/{id}/result with a token. The server never issues public
download links that work without signing in — on purpose, so your video doesn't leak through chat history.
Director rules
MCP doesn't give out the rules of the Slicer AI director — not in the brief, not in the prompts, not in tool results. When the assistant edits by itself, the rules are applied to its plan on the server, and the assistant gets fixes and issues with the plan (Your own director). A whole montage by these rules is done by our AI director — paid, at the montage price.
App-only for now
- buying token packs;
- connecting social accounts and publishing to TikTok;
- background removal, voice isolation, AI avatar, AI effects from a prompt;
- adding files to a project that already exists: pass all the materials — videos, photos, music — to
create_projectat once, and the assistant then places them in the video as fragments, inserts, music or voice; - the Chat with the AI director.
These features will come in later versions — follow the changelog.
Compatibility
The server supports MCP protocol revision 2026-07-28 and the earlier 2025-11-25, 2025-06-18 and 2025-03-26, over the sessionless Streamable HTTP transport. The legacy HTTP+SSE transport isn't supported. If a client can't connect, test the server in MCP Inspector (Connecting, the "Others" tab).