FastGPT 4.17.3 puts MCP application tool calls and external channel messages into the same team-level requests-per-minute limit as chat. Over the limit, the response carries the remaining allowance and a retry delay, and external channel messages are counted after deduplication.
For teams running automation, that single sentence changes capacity planning.
One allowance, several traffic types
The release brings MCP and external-channel requests into the team QPM check. Inventory the actual entry points used by each job, since a tool service, an application call and an upstream model request have distinct paths and may encounter different limits.
The practical consequence is that load testing has to estimate both traffic types together. Testing only one side produces a peak estimate that the production system will exceed, usually during a scheduled batch.
Response fields are part of the design
Limit responses return the remaining quota and a wait time. Both are meant to be consumed: a caller that reads the wait time can back off on schedule instead of guessing an interval.
Write the backoff into the caller’s retry configuration rather than the platform. Platform-side backoff makes the decision for every caller at once, and retries then cluster at the same moment after a peak.
Deduplication before counting
External channel messages are deduplicated before being counted, because platform-side retries can deliver the same message more than once. Counting after deduplication keeps one message from consuming quota twice, and keeps reporting aligned with actual traffic.
The same release also lists two fixes in this area: recursive references in MCP tool input schemas no longer break serialisation, and requests for unknown MCP SSE sessions now return an explicit status instead of waiting indefinitely.
Three tests worth running
Run one round of conversation traffic only, one round of MCP calls only, and one round with both. The first two give per-type peaks; the third confirms whether the combined load fits the allowance. When the third round fails, the first two identify which traffic type caused it.
Sources
Quota behaviour, vector recall limits and MCP fixes: FastGPT v4.17.3 release notes .
Separate quota from capacity
Three limits are often treated as one: the team’s chat requests per minute, the sandbox instance allowance, and the commercial edition’s vector retrieval concurrency. They are independent settings, and each governs a different stage of a request.
Record the current value and the observed peak for each, separately. When the system slows down, the first question is which of the three is being hit, and that question cannot be answered from a single combined number.
The vector retrieval concurrency setting accepts 0 to 1000 with 0 meaning unlimited, and a positive value applies across all service instances and vector databases. It has no equivalent in the open-source edition, so documentation shared between editions needs that distinction stated.
Backoff belongs in the caller
A retry policy written on the caller side can be tuned per job; a policy written on the platform applies to every caller at once. The second is easier to deploy and harder to reason about during an incident.
Three parameters matter: the maximum number of attempts, the interval between them, and whether retries count against the same allowance. That last point is easy to miss: retries are requests, and requests consume quota.
Record the chosen parameters next to the job definition. When the quota dependency is unavailable, the release notes describe logging the issue and skipping that check; callers should still handle failures and limits returned by other services in the path. During a peak, the record answers whether a spike came from new traffic or from a job retrying its way through a limit.
Watch two counters, not one
Quota consumption and latency are separate signals. A quota running near its ceiling with steady latency suggests the workload is simply large; a quota below its ceiling with rising latency points at something else, such as a slow upstream service.
Recording both per interval makes the two distinguishable. A single counter forces every incident to start from a guess about which one moved.
The same record answers the quarterly question about whether the allowance should change, and gives that answer in numbers rather than in impressions from the support channel.
