Olla now applies its proxy limits to Anthropic-compatible requests as well. Sessions using the same network address share a request allowance, which can affect parallel agents and automated jobs. The release also corrects request counting and the classification of backend failures, changing the metrics generated by existing installations after an upgrade.
Olla, a proxy that distributes language-model requests across local and remote servers, has extended its rate controls. According to its November 1 release, Anthropic-compatible messaging, token counting and model listing now fall under the proxy’s existing limits. The routes draw on a shared allowance for each internet protocol address. Parallel agent sessions and automated jobs behind the same address can therefore encounter rate limiting.[1]
Defaults allow 100 requests a minute for each address, with a burst allowance of 50, alongside a global ceiling of 1,000 a minute. Setting the address-specific limit to zero leaves the global control operating unless that is disabled separately. Olla also ends retries after any response content reaches the client, preventing a partially delivered request from being replayed to another backend.[1]
The update corrects a double count affecting requests that named a model in the preceding release. Each backend attempt now contributes one request, and backend server errors count as failures rather than successes. Those corrections affect totals, success percentages, failure counts and related alerts after an upgrade. A configuration check can inspect settings, model definitions and provider profiles without launching the server, reporting failure when validation fails.[1]
