Understand MCP work units
Estimate AI work, inspect reservations and avoid duplicate mutations.
Read as MarkdownWork, not messages
Free includes 15,000 units per UTC day and Pro 75,000 across the account, all connected clients and both execution modes. There is no fixed conversion from units to slides: a complex slide can cost more than several simple ones.
Manual editor work, manual exports and owner synchronization do not consume these units. Daily units reset at 00:00 UTC, with no rollover or automatic overage purchase.
What contributes to usage?
Formula version 1 combines changed scalar values with rounded-up units for:
| Work | Unit basis |
|---|---|
| Changed content | Each 256 characters |
| Returned JSON | Each 16 KiB |
| Media | Each 64 KiB |
| Processed or rendered pixels | Each 250,000 pixels |
A batch sums its effective changes; it is not billed as one operation merely because it is one tool call. Unchanged content is not charged as a rewrite, and no-op edits cost zero. These are creative-work measures, not model tokens or CPU time.
Inspect and estimate
Agents should call get_usage to inspect the plan, confirmed/reserved/available units and reset time. estimate_commands previews a remote batch without applying it or reserving balance.
Small status receipts, budget queries, navigation and sharing controls consume no creative units, but still have technical rate limits. Reading a large result again can incur another read charge.
Reservations and retries
Reads and renders can reserve a bounded amount before settling actual work. Uncertain outcomes may retain a reservation instead of assuming the work failed. Identical accepted mutation retries do not charge twice.
After a timeout, inspect get_request_status before retrying; use includeResult:false for a small status receipt. Keep the same request identity and inputs when retrying the same mutation.
When daily work is exhausted, wait for reset or continue manually. Storage limits, rate limits, export bounds and editor backpressure are separate constraints.