# Spending limits for assistants (/ai/spending-limits)

An assistant can only spend what you allow, and never without saying the price first. Every billed call goes through
these checks, in this order; the first one that fails refuses the call and **nothing is billed**.

| Check | What it stops |
|---|---|
| 1. **Access** | A read-only connection or key, a member when the workspace keeps members read-only. |
| 2. **`max_cost_usd`** | Any call that could cost more than the assistant agreed to pay for it. |
| 3. **Per-call limit** | Any single call above the connection's or key's limit ($0.50 by default). |
| 4. **Rate** | More than 30 billed calls a minute. |
| 5. **Balance and daily safety limit** | Running past the workspace's balance or its $25 daily safety limit. |
| 6. **Daily limit** | The connection or key spending more than its daily limit ($3 by default). |

## max_cost_usd

Every billed tool requires `max_cost_usd`: the most the assistant agrees to spend on that call. Each tool's
description states its typical price and its maximum, and the minimum `max_cost_usd` to pass. If the call could cost
more than `max_cost_usd`, PeekRank refuses it and says what it would cost:

> Not run: this call can cost up to $X (typically ≈ $Y), above your max_cost_usd of $Z. Call it again with
> max_cost_usd ≥ N to run it.

So a low `max_cost_usd` works as a free quote. Good assistants check the price with you before raising it.

## Per-call and daily limits

| | Sign-in connections | API keys |
|---|---|---|
| Max per call | $0.50 by default | Set per key |
| Max per day | $3 per connection by default | Set per key |
| Set by | Owners and admins, for every connection of the workspace | Owners and admins, on each key |
| Where | **Settings → Developers & AI → Connected apps** | **Settings → Developers & AI → API keys** → **Edit** |

Daily limits reset at 00:00 UTC. A call is refused if what the connection already spent today plus what the call can
cost goes over the limit.

## Know what's left

Ask your assistant to call `get_budget`: it returns the workspace's free data and balance, and what this connection
may still spend today. Spending through assistants shows on **Data & billing** as **Via MCP**, per connection and key.
See [See what you spent](/billing/spending/).

## Refusals and what to do

| The assistant says | What to do |
|---|---|
| This connection is read-only | Reconnect PeekRank and choose **Read + run billed tools**, or give the API key billed access. |
| The workspace lets members use PeekRank read-only | An owner or admin can turn on **Members can use billed tools**. |
| Not run: this call can cost up to … above your max_cost_usd | Agree to the price and call again with a higher `max_cost_usd`. |
| Not run: … allows at most … per billed call | An owner or admin can raise the per-call limit. |
| Too many billed calls in the last minute | Wait a minute. |
| … free data is used up / Your balance is used up | [Top up](/billing/top-up/) or wait for your free data to refill on the 1st. |
| Today's data limit … is reached | The workspace's daily safety limit: it resets at 00:00 UTC. |
| Not run: … has billed … of its … daily MCP limit today | The connection's daily limit: it resets at 00:00 UTC, or an owner or admin can raise it. |

## Good to know

- Free tools are never limited by money: they only read what PeekRank already stored.
- ChatGPT asks you to confirm every billed tool before it runs; other clients ask according to their own settings.
- Assistants never pay: top-ups are always done by a person, on Stripe's payment page.