Limits and quotas
This page documents the limits Mate enforces on jobs and tenants. Most are configurable in your workspace config; a few are fixed by your plan. Mate has no daily job cap, no job-chain depth limit and no built-in spend budget.
Concurrency
Section titled “Concurrency”| Scope | Config key | Default |
|---|---|---|
| Per-tenant concurrent jobs | Plan capacity (number of “mates” on your plan) | Depends on plan |
| Per-project job cap | policies.concurrency (map of project path to cap) |
Unrestricted |
Your plan determines how many jobs can run simultaneously across your entire tenant. (Without a subscription, as on-prem, the operator-set default of 2 applies.) On any single issue or MR, only one job runs at a time regardless of these caps; further events for it wait in the queue. policies.concurrency adds optional caps for individual projects, for example to keep agents from conflicting in one repository while another runs freely:
policies: concurrency: group/repoA: 1 group/repoB: 5Projects not listed are unrestricted (bounded only by your mates). When a project is at its cap, new matching events for it are queued and do not take up tenant capacity while they wait. When the per-tenant pending-queue caps fill up, excess events are rejected (see Queue backpressure below).
A plain number (concurrency: 3) is no longer accepted: policies.concurrency is a map of project path to cap, each at least 1. Remove the scalar to use all your mates.
Job timeout
Section titled “Job timeout”| Config key | Default | Scope |
|---|---|---|
agents.<name>.timeout |
Inherits policies.default_timeout |
Per-agent |
policies.default_timeout |
1 hour (when neither timeout nor default_timeout is set) |
Per-workspace |
A job that runs longer than its timeout is canceled with reason timeout and ends as failed. The status note on GitLab names the limit and how to raise it (“Stopped: reached the 30m time limit. Raise it with timeout: on the agent or policies.default_timeout in the workspace config.”). Set timeout: on an individual agent to override the default for that agent:
agents: slow-refactor: backend: claude-code model: anthropic/claude-opus-4-7 timeout: 60m # ...No-progress watchdog
Section titled “No-progress watchdog”A job with no progress for a fixed window (20 minutes by default) is automatically canceled with reason timeout and ends as failed. Progress means any event received from the agent other than a heartbeat — a tool call, a log line, or a status transition. The status note reads “Stopped before finishing” with the reason (for example no progress for 20m0s).
This watchdog catches stuck agents that are still technically running but have made no progress. It cannot be disabled; the window is not configurable per agent or workspace (on-prem operators can change it with MATE_NOPROGRESS_IDLE). Two sibling checks use the same mechanism: a job that never reports in within 20 minutes of being submitted, and a running job that stops sending heartbeats for 3 minutes, are stopped the same way.
Tool calls to GitLab can legitimately take long. A single call may run for up to about 15 minutes; wait_pipeline_finish and wait_pipeline_change are capped at 840 seconds per call and the agent re-issues them to wait longer.
Cooldown between repeat triggers
Section titled “Cooldown between repeat triggers”policies.cooldown sets the minimum time that must elapse between consecutive job dispatches on the same repository. This prevents a fast-looping comment thread from spawning a cascade of jobs:
policies: cooldown: 5m # no more than one job every 5 minutes per repoDefault is no cooldown (each matching event dispatches immediately). The window is measured from the previous job’s start, not its end. While a cooldown is set, a repository also runs only one job at a time.
LLM spend
Section titled “LLM spend”Mate does not enforce an LLM spend cap, and the former policies.budget key no longer exists. Mate is BYOK: set budgets and rate limits with your LLM provider (for example a credit limit on the key you give Mate).
The one spend-related setting that remains is operator-level, for on-prem installs that mint per-job OpenRouter sub-keys from a provisioning key: openrouter_default_limit_usd in the tenants file caps each sub-key in USD. It is not a workspace-config key.
Queue backpressure
Section titled “Queue backpressure”Mate queues incoming events up to a per-tenant limit (1024 pending jobs by default) and a global limit across all tenants (1,048,576). When either is full, new webhook deliveries are rejected with a 503 response. GitLab interprets repeated 503 responses as a degraded webhook and may temporarily disable it.
In practice, this limit is only reached under sustained high event volume with slow or stuck jobs. Resolving the underlying cause (jobs completing, cooldown configured, concurrency tuned) clears the backlog automatically.
Container resources
Section titled “Container resources”Container resources are determined by your plan and are not configurable per agent.
| Plan | vCPU | Memory |
|---|---|---|
| Shared | 1 | 2 GiB |
| Dedicated | 1 | 2 GiB |
On-prem deployments run job containers on the customer’s own hardware with the same per-container ceiling by default (1 vCPU, 2 GiB); the on-prem enterprise tier raises it to 2 vCPU and 4 GiB.
Context size limits
Section titled “Context size limits”| What | Limit |
|---|---|
| Text of the @-mentioning comment, passed to the agent as the triggering excerpt | 32,000 characters; longer text is cut, and the agent is told it was truncated and can fetch the rest through GitLab. |
Repository CLAUDE.md (else AGENTS.md) added to the claude-code system prompt |
512 KiB; a longer file is cut with a note telling the agent to read the rest from the checkout. |