max_spend_usd is optional on Sume MCP: send it on every paid call

On Sume's hosted MCP server, max_spend_usd is optional, and Sume enforces it only when you provide it. A coding agent that omits it has no per-call ceiling beyond wallet admission, so tell your agent to send it on every paid call, together with dry_run for a preview and an idempotency_key, which is required.

The gates, one by one

The gates table in the MCP docs is short, and each row is different. idempotency_key is required on write and paid tools, and it is a transport and de-duplication key, not human approval. dry_run=true is an optional admission and cost preview that does not submit the job. max_spend_usd is optional and enforced only when supplied. The legacy allow_write and allow_paid flags are accepted for compatibility, are not required, and cannot bypass a missing mcp:write scope.

What really stops a paid call

The docs are direct about the spend gate: spend is wallet and admission, and there is no mcp:paid scope. Under OAuth with only mcp:read, the write and paid tools are hidden and return insufficient_scope. With mcp:write, or with an API key, the full tool set is available, and a paid submit still needs its key and admission.

Put the habit in the agent's instructions

That makes the optional gates a client-side habit. The docs recommend generation_admission_preview or dry_run before expensive bursts, and say that a normal single create does not need them. So write the habit into your agent's standing instructions, and keep the numbers small.

A standing instruction

A standing instruction like the one below works in a project file that your agent reads at the start of a session. Edit the amount to match your own limit.

Sume paid tools (generate_video, generate_image, music_create, tts_create,
avatars_create, avatar-videos_create):
1. Call tools_schema for the tool first.
2. Run it once with dry_run=true and show me the cost.
3. Submit only after I say go, with a fresh idempotency_key and
   max_spend_usd set to the previewed cost plus 20 percent.
4. If jobs_wait returns wait_slice_expired, call jobs_wait again on the
   same job id. Never call the create tool a second time.

Why the last line is there

Step 4 matters as much as the cap. The remote jobs_wait hold is 50 seconds by default, with a cap of 55, and a slice that expires is not a failure. The job is still running, so a second create would pay twice.

Short version

Set the ceiling on every paid request, preview anything large, and keep the idempotency key stable for each retry.