---
title: "Usage limits"
description: "Set shared token and per-person reply limits, bound each response, and choose how context is carried through long conversations."
image: "https://mazegraph.com/docs/og.png"
---

> Documentation Index
> Fetch the complete documentation index at: https://mazegraph.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Usage limits

MazeGraph’s usage controls limit provider work at different levels: each reply has a step and response-length limit, each person has a message pace, and the whole site shares a daily token allowance. An administrator can tune these in **MazeGraph → Settings → Limits and privacy**. The controls help manage activity, but they are not a money limit set by your AI provider.

## Limits at a glance

| Control | Scope | Default | What it limits |
| --- | --- | --- | --- |
| **Steps per reply** | One reply | 8 model calls | The number of model calls in a reply, including tool rounds. The final call cannot run tools. |
| **Longest answer** | One model call | 0, which uses a 4,096-token output limit | How many output tokens the model may generate in each call. |
| **Daily allowance** | Whole site, per UTC day | 500,000 tokens | The site's combined token use across model calls. |
| **Messages per person per hour** | One WordPress user | 60 replies started | How many replies one person can start in a rolling hour. |
| **Long conversations** | Context sent to a model | Recent messages only | How earlier messages and tool results are included as a conversation grows. |

## Bound each reply

**Steps per reply** counts model calls, not just the message a person typed. A reply that asks the model to use a tool takes a model call to choose the tool and another call after the tool result. The last allowed call is reserved for the model’s answer and cannot start another tool call. The setting accepts 1–20 calls and defaults to 8.

Some agents have a lower step limit of their own. MazeGraph uses the lower of the site limit and the selected agent’s limit, so a reply can stop before it reaches the site maximum. See [Agents](/docs/tools/agents/) for the agent catalogue. If a task needs more than the available steps, break it into smaller requests instead of raising every limit automatically.

**Longest answer** is the maximum output for each model call, rather than a single allowance for an entire multi-step reply. The default value, 0, means MazeGraph applies a 4,096-token output limit; it does not mean unlimited. A smaller value can make an answer or intermediate model response shorter. The setting accepts values up to 200,000 tokens, but larger responses can use more of the shared daily allowance and may affect provider charges.

## Set a site allowance and message pace

**Daily allowance** is shared by everyone on the WordPress site. The default is 500,000 tokens each day, counted from midnight UTC. Set it to 0 to remove MazeGraph’s daily cap. When the allowance is reached—or when a new model call cannot safely fit within what remains—MazeGraph blocks further model work until the allowance resets or an administrator changes it.

Before each model call, MazeGraph reserves an estimate against the shared allowance and reconciles it with provider-reported usage when that information is available. This lets simultaneous requests share one site-wide limit. If a provider request fails and its actual usage is unknown, MazeGraph keeps the reservation to avoid assuming that the call cost nothing.

**Messages per person per hour** limits how many replies a single WordPress user can start during the previous hour. It defaults to 60; set it to 0 for no per-person limit. The pace is per person, while the daily allowance is shared across the site.

Provider billing is separate from these controls. MazeGraph can display recorded token totals and estimated costs, but your provider or connector determines the actual charge. Review [Usage in settings](/docs/settings/usage/) and [Understand usage totals](/docs/activity-and-usage/understand-usage-totals/) when you need to compare activity or estimates.

## Choose context for long conversations

The **Long conversations** setting controls what MazeGraph includes in model requests as history grows. It does not determine how long the saved session is retained; see [Limits and privacy settings](/docs/settings/limits-and-privacy/) for retention controls.

| Option | What MazeGraph sends |
| --- | --- |
| **Recent messages only** | Keeps the newest conversation messages that fit the context window. Older messages may no longer be included in later model calls, but this option does not delete them from the saved session. |
| **Drop old tool results** | Keeps recent tool results and replaces older successful results with a short note. The assistant can run a tool again if it needs that information. |
| **Summarize older messages** | Creates and reuses a summary of earlier conversation when the history grows. Each summary needs one additional model call and uses tokens from the site's daily allowance. |

Choose recent messages for the simplest behavior. Drop old tool results when repeated tool output is making a conversation heavy. Summaries can help retain earlier context during a long task, with the tradeoff of extra model calls and token use.

## Adjust the limits

1. In WordPress, open **MazeGraph → Settings → Limits and privacy**.
2. Review **Reply limits**, **Usage limits**, and **Conversation context**.
3. Change only the values that address the issue, then select **Save changes**.
4. Review the [Usage report](/docs/settings/usage/) to see how site activity changes.

If replies stop at the daily allowance, an administrator can raise the cap or wait for midnight UTC. If one person reaches the hourly limit, they can wait before starting another reply or an administrator can adjust the pace. For a reply that ends at its step or length limit, see whether the task can be split before increasing the relevant setting.

## Diagnose a reply that stops

| What happened | Likely cause | What to do | What to review next |
| --- | --- | --- | --- |
| A reply stops after several tool rounds. | It reached the site step limit or the selected agent’s lower limit. | Reduce the task to fewer steps, or have an administrator review the site and agent limits. | [Agents](/docs/tools/agents/) and [How MazeGraph decides what may happen](/docs/workspace-controls/how-mazegraph-decides-what-may-happen/). |
| New model work is refused for everyone. | The shared daily allowance is exhausted or cannot reserve enough for another call. | Wait for midnight UTC or ask an administrator to review **Daily allowance**. | [Usage in settings](/docs/settings/usage/). |
| One person cannot start another reply. | That person has reached the hourly message pace. | Wait, then try again; an administrator can review **Messages per person per hour** if the limit is too restrictive. | [Availability and role access](/docs/workspace-controls/availability-and-role-access/) for other site-wide controls. |
| An answer ends sooner than expected. | The model call may have reached **Longest answer** or the reply step limit. | Check the end-of-reply notice and ask an administrator to review both settings. | [Limits and privacy settings](/docs/settings/limits-and-privacy/). |

For a broader review of refusals and approval requirements, see [Policy Administration and Safeguards](/docs/policy-and-safeguards/). For details about session and activity-log retention, see [Activity and Usage](/docs/activity-and-usage/).

Source: https://mazegraph.com/docs/workspace-controls/usage-limits/index.mdx
