---
key: api-limits
title: Rate limits, daily caps and other limits on a credential
audience: developer
section: api/authentication
routes: []
since: 0.0.1
updated: 0.0.1
status: published
summary: The per-minute limits, the optional daily cap on changes, working hours, address allowlists, expiry and branch limits that can apply to an API key, connected app or AI teammate.
keywords: [rate limit, 429, throttle, daily limit, quota, ip allowlist, expiry, rotation, working hours, branch]
related: [api-errors, api-conventions, choose-an-api-key-or-a-connected-app, run-an-ai-teammate]
---

What a credential *may do* is its grant ([[api-scopes-reference]]). The
limits below sit beside the grant and say *how much, from where and when*.
The organization sets them on the credential; nothing here is decided by Nova
on its own.

## Per-minute limits

| Limit | Amount | Applies to |
|---|---|---|
| Per credential, standard | 600 requests a minute | Every API key, connected app and AI teammate |
| Per credential, high volume | 3,000 requests a minute | Credentials an admin set to **High volume** |
| Per token | 600 requests a minute | Each access token or API key |
| Per organization | 3,000 requests a minute | All software in one organization together |

Going over answers:

```json
{"error": {"code": "rate_limit_exceeded",
           "message": "Too many requests. Please try again later.",
           "details": {"limit": 600, "window_seconds": 60,
                       "retry_after_seconds": 60, "count": 601}}}
```

Wait `retry_after_seconds`, then carry on. Spread bulk work out rather than
bursting.

Over MCP, every message counts once against the per-minute limits, whatever
it does: `tools/list`, `tools/call`, or any other. The Nova operation a tool
runs is part of that message and is not counted again. A JSON-RPC batch
counts once per message in it, and a batch that would go over the limit is
refused whole, before any of it runs. Going over answers HTTP `429` with the
`rate_limit_exceeded` body above, never `401`, so do not sign in again; wait
and retry.

## Daily cap on changes

An admin can cap how many **changes** an API key or an AI teammate makes in
a day (1 to 100,000). The cap belongs to that credential: a connector that
connects as an AI teammate is counted against its own credential, not the
teammate's. Reads (`GET`) never count. Over MCP every message is a `POST`,
so a tool call is judged by the operation it runs: listing tools and calling a
read tool never count, and a tool that changes something counts once. Once the cap is reached every change
answers `429 rate_limit_exceeded` with `window_seconds: 86400` until the
24-hour window that began with the first counted change runs out. The
person accountable for the credential gets one notification in Nova the
first time it happens that day.

## Working hours (AI teammates)

An AI teammate can be given working hours. Outside them it can still read,
but a change answers `403` with `detail.error: agent_outside_working_hours`
and `detail.next_window_starts_at`. Over MCP the same rule holds per tool: an
off-shift teammate can list tools and call read tools, and a tool that changes
something returns that error as its result. See [[run-an-ai-teammate]].

## Where it may be used from

A credential can be limited to a list of addresses or ranges
(`203.0.113.4`, `203.0.113.0/24`). A call from anywhere else answers
`403` with the message *"This credential can only be used from an address its
owner allowed."* Empty means anywhere.

## Branches

A credential can be limited to some of the organization's branches. It then
sees and changes only work (jobs, quotes, invoices, payments, bookings, time)
that belongs to those branches, exactly as a person limited to those branches
would.

## Expiry and replacing a key

- An API key has an expiry (30 days to 2 years, or never). An expired key
  answers `401`.
- Replacing a key issues a new one and keeps the old one working for an
  overlap you choose (straight away, 1 hour, 1 day or 1 week), so you can
  update your software without an outage.
- Pausing or retiring an AI teammate revokes its tokens immediately.

## Paid actions

Texts, emails and other actions that cost money are bounded by the
organization's daily budget. When it is spent, the call answers
`429 spend_deferred` with `details.retry_at` (the next day). See
[[api-errors]].