Getting started for AI agents
An AI agent can work with Crumbcast in two ways:
- Through the MCP server, as a tool user: Claude Code, Cursor or your own agent finds, imports, captions, cuts and organises a project's videos, puts them in playlists, makes links to share them, and runs its live streams, while you talk to it.
- By writing code against the API, as a coding assistant: it reads the API's spec and writes the uploads, webhooks and signed playback your product needs.
Either way, it acts with a project's API token, held to the same rules as any other code: the token's project only, at the token's role.
Use Crumbcast from an agent (MCP)
1. Make a token
In the project's Settings → API tokens, make a token for the agent, with the least role
it needs: viewer to look without changing anything, editor to import, caption, cut,
organise, share and go live.
2. Add the server
The server is at https://your-crumbcast/api/v1/mcp (Streamable HTTP), with the token as
its bearer token. In Claude Code:
claude mcp add --transport http crumbcast https://your-crumbcast/api/v1/mcp \
--header "Authorization: Bearer cc_…"
Other clients take the same address and header in their JSON settings; see AI agents (MCP).
3. Ask it something
Start with something that only reads, to check it's connected:
What's in my Crumbcast project? Which videos are still processing?
Then give it work:
Import https://example.com/talk.mp4, caption it, and suggest chapters for me to look over.
Processing runs in the background: the agent starts it, then checks back on the video. The tools it has, and the role each needs, are in the guide.
Have an agent write code for the API
Give your coding assistant the API's OpenAPI spec, which lists every endpoint a token can call, with its fields, answers and errors:
https://your-crumbcast/api/v1/openapi.public.json
and point it at the guide for the task, so it follows the steps Crumbcast expects:
- Uploading a video: a file in parts, straight to storage
- Webhooks: hearing when a video is ready, and checking each call is Crumbcast's
- Signed playback: playback tokens made on your backend
If your code is JavaScript, TypeScript, Python or PHP, have it use the JavaScript, Python or PHP SDK: their methods and types are the spec's, and uploads in parts and webhook checks are already done right.
Keep the token out of the code it writes: read it from the environment
(CRUMBCAST_TOKEN), on your servers only.
What an agent can't do
Like any token, an agent's can't reach your account, organisation, storages or billing, nor
manage the project's members or tokens. What the API refuses, the agent is told why
(403: …, 402: …), so it can tell you.