Comfy Agent Auto mode burns credits per token: how to stay capped

Comfy Agent charges tokens from the same Comfy Credits pool you already use for generations, and you need a positive balance to keep running it, though not necessarily a subscription. In Auto mode it runs workflows without asking for each run. The page I read documents a Stop control for the current response and the balance requirement, and I found no per-chat spending limit on it, so the credit balance is your ceiling. On Sume, the matching ceiling is a number you set on each run.

Rates and modes are from Comfy's Agent documentation, read on 2026-10-03; Sume's side is in Agent Completions and MCP tools and gates.

What does Comfy charge?

The page lists per-token rates for the model it names. These are credits per 1,000 tokens.

What do Ask and Auto change?

Ask requests approval before each workflow run; Auto runs workflows without asking each time. Token cost accrues while the Agent plans and iterates, so Auto mode combines two meters: tokens for thinking and credits for each generation it launches. Select Stop in the message box to end the current response.

How do I cap the Sume side when Comfy Agent calls out?

If a Comfy workflow reaches Sume, for example through a custom node calling the API, the credit pool above does not apply. Sume has its own balance and admission. Send generation_spend_cap_usd on every Agent Completion, or max_spend_usd on every paid MCP call, because Sume enforces the MCP cap only when you send it. Use dry_run=true to preview before a burst, and generation_admission_preview before a wave.

A balance shortfall fails at submit with 402 insufficient_credits before provider work starts, so an agent loop can be told to stop on that code instead of retrying.

What is a safe default?

One practical habit: check the balance before leaving an Auto session unattended. Because cache hits are priced far below fresh input at 0.106 credits per 1K tokens against 1.055, long sessions that reuse context cost less per token than many short ones that rebuild it, but the page still gives no total limit, so the balance is the only brake.

• Start in Ask mode on a new workflow, switch to Auto only after one clean run.
• Keep a small positive balance on the account you hand to an unattended agent.
• Cap each Sume call in dollars, so Comfy's token meter and Sume's generation meter are bounded separately.
• Stop on 402, and read retry-after on 429 rather than looping.