Skip to main content

AI agents (MCP)

Crumbcast is an MCP server, so AI agents and coding assistants (Claude Code, Cursor, your own agent) can work with a project's videos and live streams: find a video, import one, caption it, cut it, read its numbers, go live.

The server is at https://your-crumbcast/api/v1/mcp (Streamable HTTP). It signs in with a project's API token, sent as the bearer token, like the API's requests, and it's held to the same rules: the token's project only, at the token's role. Give an agent a viewer token to let it look without changing anything.

Connecting​

Claude Code:

claude mcp add --transport http crumbcast https://your-crumbcast/api/v1/mcp \
--header "Authorization: Bearer cc_…"

Clients configured with JSON (Cursor, Claude Desktop through mcp-remote, and others):

{
"mcpServers": {
"crumbcast": {
"url": "https://your-crumbcast/api/v1/mcp",
"headers": { "Authorization": "Bearer cc_…" }
}
}
}

Clients that only sign in with OAuth, such as claude.ai's connectors, can't connect yet.

Tools​

ToolDoesRole
project_overviewThe project at a glance: videos, storage, views, live streamsviewer
list_videosThe project's videos, by title words, folder (or none), metadata, still processing or failed, a page at a timeviewer
get_videoA video's formats, captions, chapters, links, and what may be done to it nowviewer
import_videoA new video from a public URL, with a workflow and metadata if you likeeditor
update_videoIts title, description, folder, labels, playback policy and password, chapters, metadata, ad tageditor
delete_videoDeletes a video: into the project's trash if it keeps one, else for goodeditor
list_trash, restore_videoThe deleted videos the trash keeps, and bringing one back as it wasviewer, editor
move_videos, label_videosMoves videos into a folder, or adds a label to them or takes it offeditor
create_folder, create_labelA new folder or label to sort videos intoeditor
generate_captionsCaptions from the video's speecheditor
get_transcriptWhat's said in a video, from its captionsviewer
update_captionsA caption track's cues, correctededitor
suggest_chaptersChapters read from its captions, to look over before savingeditor
clip_videoA new video of the parts kepteditor
trim_videoCuts the video itself to the parts kepteditor
retry_encodingEncodes a failed format's failed parts againeditor
set_thumbnailA frame of the video as its thumbnaileditor
run_workflowRuns one of the project's workflows on a video already ineditor
list_folders, list_labels, list_workflows, list_ad_tagsWhat a video can be sorted into, have done to it, or playviewer
list_playlists, create_playlist, update_playlistPlaylists, and the videos in them in play orderviewer, editor
share_linkA watch link with a viewer link's token in it, for a signed or password video, live stream or playlisteditor
get_analyticsViews and watch time, of the project or one videoviewer
list_live_streams, get_live_streamLive streams, whether they're live, and where to push toviewer
create_live_stream, update_live_streamA live stream and its RTMP, SRT and WHIP ingest; its title, recording, latency and policyeditor
delete_live_streamDeletes a live stream; its recordings stayeditor

Each tool describes its answer (its output schema), so an agent knows what it gets back before it calls. list_videos and list_trash answer a page at a time: has_more says whether there's another, and next_cursor, passed as after with the same filters and order, asks for it.

Processing runs in the background, as it does through the API: a tool that starts it answers at once, and the agent checks on it with get_video. What the API refuses, a tool does too, with the same reason (403: …, 402: …), for the agent to read.