Security

Connecting an LLM to an account that can create clip projects, create and remove events, and update destinations carries the usual agent risks.

Same-access principle

The server operates as the signed-in user. Grant it only to people who should already have that access.

Keep human confirmation on

Leave it enabled and review write tool calls before approving them.

Prompt injection

If an agent processes untrusted content and that content contains instructions, review planned actions before approving, and don't auto-approve write tools when untrusted input is in context.

Removals are non-destructive externally

toggle_event_channel_off and remove_event only change Restream state; they never delete or modify anything on the destination platforms.

Verify the endpoint

Only connect to the official https://mcp.restream.io/mcp.

Credential boundary

The server never accepts passwords or RTMP/SRT credentials and never returns stream keys. Handle those in the dashboard or REST API.

Current limitations

• No password, RTMP/SRT credential entry, or stream-key retrieval.
• Not yet listed in the Claude or ChatGPT connector directories — add the server manually with the endpoint above.