Limit an agent's paid MCP calls with script_run max_paid_calls

script_run on Sume's hosted MCP runs a short JavaScript program that calls other Sume tools in a loop, in parallel, or with conditions, and returns one value. Its limits are timeout_seconds (5 to 55), max_calls, and max_paid_calls. It is a way to bound how many paid calls one agent turn can make, but it is not a dollar cap.

What the docs say

The tools and gates page recommends script_run when a turn needs three or more independent calls of the same shape, such as one tts_create per sentence or one generate_image per scene. Inside the script, await sume.call(name, arguments) runs any listed tool with the same gates, redaction, and errors as a direct call.

Why it helps with a cheap, fast agent model

Fast models such as Claude Haiku 5.5, which Anthropic's page (read 2026-10-08) pitches for subagents and high-volume work, will happily take many small steps. Each step pays prompt tokens again. Moving a fan-out of ten identical tool calls into one script means the model pays for one call and one returned value instead of ten results in its context.

The count limits then bound the blast radius of a bad plan. If the model decides to loop over 400 scenes by mistake, max_paid_calls stops the script at the number you set.

What it does not do

There is no dollar field in the list above. For a dollar ceiling, use the gates the docs name elsewhere: max_spend_usd on a paid tool, which Sume enforces only when you provide it, and dry_run=true for an estimate that does not submit. Wallet balance and generation admission still apply on every paid call.

Also note the scope rule. Under OAuth mcp:read the write and paid tools are hidden, and the docs say there is no separate paid scope; you need mcp:write or an API key.

A practical pattern

The points that matter here, in the order you will hit them:

• Call generation_admission_preview or a dry_run first for any burst.
• Set max_paid_calls to the number of items you intend, not a round large number.
• Pass a stable idempotency_key per item so a retried script does not double-submit.
• Wait on the returned jobs[] with jobs_wait rather than re-running the script.

Choosing the numbers

Suppose a script renders 12 scenes, then waits for the jobs. That is 12 paid creates. Setting max_paid_calls to 12 means a thirteenth paid call stops the script. max_calls has to cover the paid creates plus any reads, so a value of 14 would allow two reads. The timeout is separate: 5 to 55 seconds, so long generation work is not waited on inside the script. The docs return child jobs[] so you can use jobs_wait afterward.