MCP server

ReelMoney runs a Model Context Protocol server at https://reel.money/mcp. Point Claude, Cursor, or any MCP client at it and the model gets 18 tools that do the same work as the REST API, calling the same code underneath, so the two surfaces can never drift.

There is no OAuth flow yet. The key goes in the header, which every client below supports.

Connect a client

Claude Code, one line in a terminal:

claude mcp add reelmoney --transport http https://reel.money/mcp --header "Authorization: Bearer rm_your_key"

Claude Desktop, Cursor and anything else that reads a JSON config:

{
  "mcpServers": {
    "reelmoney": {
      "type": "http",
      "url": "https://reel.money/mcp",
      "headers": {
        "Authorization": "Bearer rm_your_key"
      }
    }
  }
}

A raw check with curl, useful when a client is quiet about why it failed:

curl -X POST https://reel.money/mcp \
  -H "Authorization: Bearer rm_your_key" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

The tools

Every tool carries a title and the standard annotations, so a client can tell a read from a write before it calls anything.

Rules a client should know

• A missing scope does not hide a tool. Every key sees all 18 tools. A tool the key lacks the scope for answers at call time with a structured missing_scope error that names the scope and how to add it. Predictable beats surprising.
• **brand_id picks the brand, when the key allows it.** Every brand facing tool takes an optional brand_id argument. Omit it and the tool works on the key's own brand. A key with the brands:all scope may pass any brand id the account owns (get_account lists them); without the scope a foreign brand_id answers a structured missing_scope error, and an id the account does not own answers not_found.
• Two tools are marked destructive. schedule_post and cancel_post carry destructiveHint and _meta: {"anthropic/requiresUserInteraction": true}. Publishing to an audience and pulling a post back are decisions a person should confirm. Treat the prompt as the point of the annotation, not as a formality.
• GET /mcp answers 405. The server is stateless, so there is no session to attach a server sent event stream to. Use POST. DELETE /mcp answers 200 and does nothing, since there is no session to end.
• Bad key, 401 at the transport. Authentication happens before any protocol handling, so an invalid key answers HTTP 401 with WWW-Authenticate: Bearer realm="ReelMoney", error="invalid_token" and the usual error envelope, not a JSON-RPC error.
• Rate limits are shared with REST. The same per key ceilings apply. See authentication.

Result size and pagination

Tool results are token budgeted, around 25k tokens in Claude Code, and a blown budget wastes a whole turn. So:

• Lists come back capped at 20 items with a note. Pass page to walk further instead of asking for everything.
• Prefer a list tool with a filter over a list tool without one. list_shorts with short_type and status beats reading the whole library.
• get_short returns full detail including slides and overlays. Call it for one record, not in a loop over a list.
• get_analytics without arguments returns account summaries; add social_data_id only when you need post by post numbers.

A working loop

1. get_account. Read the plan and how much quota is left. Stop here if there is not enough.
2. get_brand and, when you need a specific look or a media pack, list_media.
3. create_slideshow or create_video. Both answer with a job_id.
4. get_job, no faster than poll_after_seconds, until it reads succeeded or failed.
5. render_video when a caller wants an MP4 of a slideshow. Never create a second slideshow for that.
6. list_connections, then validate_post, then schedule_post. Read every verdict before you post, and every result after.
7. get_analytics a day or two later to see what it did.
8. explain_error for any code you do not recognise, instead of guessing from the message.