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

# Rate Limits

> Per-workspace request-rate and concurrent-stream limits enforced at the edge. Requests that exceed the limit get a 429 before reaching your provider.

Toolken enforces two types of limits at the gateway, before a request reaches your provider.

* **Request rate limit** - how many requests your workspace can send per minute.
* **Concurrent stream limit** - how many streaming responses your workspace can hold open at once.

<Warning>Self-serve rate limit configuration is coming soon. Limits today are set by the Toolken team. [Contact us](mailto:hello@toolken.ai) to adjust limits for your account.</Warning>

## What you see when a limit is hit

Both limits return an OpenAI-compatible error body.

```json theme={null}
{
  "error": {
    "message": "Rate limit exceeded",
    "type": "proxy_error",
    "param": null,
    "code": "rate_limited"
  }
}
```

<Table>
  <Thead>
    <Tr>
      <Th>HTTP status</Th>
      <Th>Error code</Th>
      <Th>Meaning</Th>
      <Th>What to do</Th>
    </Tr>
  </Thead>

  <Tbody>
    <Tr>
      <Td>`429`</Td>
      <Td>`rate_limited`</Td>
      <Td>Request rate exceeded</Td>
      <Td>Check the `Retry-After` header for the number of seconds to wait, then retry. Add backoff if you consistently hit this limit.</Td>
    </Tr>

    <Tr>
      <Td>`429`</Td>
      <Td>`concurrent_stream_limit`</Td>
      <Td>Too many streaming responses open at once</Td>
      <Td>Wait for open streams to complete before sending additional streaming requests. Consider reducing parallelism in agent loops.</Td>
    </Tr>
  </Tbody>
</Table>

When the request rate limit fires, the response includes a `Retry-After` header with the number of seconds until your window resets.

## How enforcement works

Both limits are checked at the edge before the request is forwarded. A request that exceeds either limit is rejected immediately with a `429`. An open streaming response counts against your concurrent limit until it finishes.

<Note>If the enforcement backend is briefly unavailable, requests are allowed through rather than blocked. Toolken prioritizes availability over hard-blocking during transient disruptions.</Note>

## Handling 429s in your code

A `429` from Toolken is safe to retry. Read `Retry-After` and wait that many seconds before the next attempt. For agent loops that fire many parallel requests, add a concurrency limit on the client side to stay within the per-workspace ceiling.

```python theme={null}
import time
import openai

def chat_with_retry(client, **kwargs):
    for attempt in range(5):
        try:
            return client.chat.completions.create(**kwargs)
        except openai.RateLimitError as e:
            retry_after = int(e.response.headers.get("Retry-After", 5))
            time.sleep(retry_after)
    raise RuntimeError("Rate limit retries exhausted")
```

## Next

<CardGroup cols={1}>
  <Card title="Observability" icon="chart-line" href="/features/observability/overview">Track request volume and spend before you hit a limit.</Card>
</CardGroup>