Usage quotas and concurrency

Cap how many tool calls can be made in a time window, and how many can run at the same time. Use quotas to stop a runaway agent or a heavy user from exhausting an MCP server or its upstream API limits.

Policy Engine only

There is no quota setting in a server's Configure sections. Quotas and concurrency limits are set only on the Usage Quotas & Concurrency card (stage 3, Request) in Govern > Policy Engine > MCP Gateway, and apply once the Policy Engine is on for the MCP Gateway.

Where to set it

Go toQuilrAI consoleGovernPolicy EngineMCP GatewayUsage Quotas & Concurrency

Edits join a shared draft and apply once you publish a revision.

Effects

Effect	Key	What it sets
Requests per minute, per hour, per day	quota.minute, quota.hour, quota.day	The number of calls allowed in each window.
Quota window	quota.window	fixed or rolling.
Quota timezone	quota.timezone	The timezone used for the quota windows.
Quota dimensions	quota.dimensions	What the counter is keyed by: tenant, user, agent, mcp, tool or group.
Quota ID	quota.id	An identifier for the quota counter.
Concurrent requests	concurrency.limit	How many calls may be in flight at the same time.
Concurrency TTL	concurrency.ttl_seconds	A time limit, in seconds, for a concurrency slot.
Concurrency dimensions	concurrency.dimensions	What the concurrency limit is keyed by, using the same dimensions as quotas.

Dimensions decide who shares a counter. user, mcp gives each person a separate allowance on each MCP server. tenant gives everyone one shared allowance.

Quotas are reserved all or nothing. A call that would cross any limit is refused rather than partially served.

Example: per-user budget on direct connections

per_user_tool_budgetrequest

runs on request·priority 500

WhenRoute kindisdirect

Then

requests per minute→60

requests per day→5,000

quota window→rolling

quota dimensions→usermcp

concurrent requests→4

Each person can make 60 calls a minute and 5,000 a day on each MCP server they reach directly, with at most 4 running at once.

More scenarios

• Protect an upstream API with a tenant-wide cap. Match MCP name and key the quota by tenant and mcp, so the whole organization stays under the provider's own rate limit.
• Limit expensive tools only. Match tool name or tool tags and key by user and tool, so a costly search or export tool has a lower budget than the rest of the server.
• Tighter limits for one group or agent. Match smart groups or agent name, for example a lower daily quota for an automated agent than for people.

Related

• Group and user rules - per group overrides that do not need the Policy Engine.
• Server access - decide who may use a server at all.
• Policy Engine overview - how cards, stages and priorities work.