AI Content Drop API Rate Limits

The limits

The MCP ceiling is deliberately higher than the REST one. A single MCP client session is not one request: connecting fires an initialize, then tools/list, resources/list and prompts/list, then one read per resource — a hundred-request burst before the client has done anything a user asked for. A limit sized for REST refuses an ordinary client mid-handshake.

Read limits are bucketed per IP. A valid agent-registered token selects the higher ceiling for that IP without unlocking account data or generation. A user API key authenticates account and generation operations but does not select the registered-agent read tier.

Reading the headers

Every response — not just a 429 — carries the standard fields:

RateLimit-Limit: 120
RateLimit-Remaining: 118
RateLimit-Reset: 47
RateLimit-Policy: 120;w=60

Self-throttle on RateLimit-Remaining rather than waiting to be refused. You can observe them with no credential at all: curl -sD - https://aicontentdrop.com/v1 -o /dev/null.

When you are refused

A 429 carries Retry-After in seconds and a JSON body naming the window. Wait that long and resume — retrying sooner extends the window rather than shortening it.

{
  "error": {
    "code": "rate_limited",
    "message": "Rate limit exceeded. Wait 31 seconds and retry.",
    "retry_after_seconds": 31,
    "documentation": "https://aicontentdrop.com/docs/rate-limits"
  }
}

Generation budgets

Generation endpoints carry their own hourly and daily budgets on top of the request rate, tied to the plan on the account. Those are about spend, not traffic: the error message names the specific window that was hit. Unlimited-model access has a separate concurrency and hourly-delay engine, described on pricing.

What is not rate limited

The static discovery documents — /llms.txt, /openapi.json, /.well-known/*, markdown twins, and /graphql/schema.graphql — are cached and served without a limit. Read them freely.