• Howler is an MIT-licensed macOS menu-bar app from developer Aries Zhou that clears messages when its chat window closes, per its README.
• Requests go directly to the AI service you configure, with no relay, according to the project's own documentation.
• Local erasure does not touch provider-side logs; The Register's report says not to assume the provider keeps no record.
• We did not run or audit the code, so every behavioral claim here is the project's own statement.
• Signal or Noise: Signal for the ephemeral-by-design pattern; the privacy payoff depends on the provider you pick.

Howler is a small open-source macOS app that answers quick AI questions and keeps no chat history, and the design holds up better than its novelty framing suggests. Messages live only in memory, requests go straight to the model service you configure, and the window clears when closed, according to the project's README and The Register's report. We did not run or audit the code. The open question is not the app but the provider behind it.

What the project says it does

Howler is a menu-bar app for macOS 14 and later, written in Swift under an MIT license, according to its repository. The README says messages and parsed-answer caches are cleared when the chat closes, and requests go to the configured service with no relay. The Register reports the same claims from the project's site and names the developer as Aries Zhou. These are the maker's statements; we have not independently tested them.

Where the data actually goes

Howler sends each question directly to whichever service you configure, so the provider's retention terms still apply. The README lists presets for OpenAI, Anthropic, Gemini, xAI, DeepSeek and OpenRouter, plus local Ollama and LM Studio. The Register's report warns not to assume the provider keeps no record of the exchange. Pointing Howler at a local model is the only configuration in which no third party sees the prompt.

What the design gets right

The architecture removes several common leak paths, if the README is accurate. It states there is no relay server, the API key is stored in the macOS Keychain rather than the chat, and the only other outbound call is a daily update check that can be disabled. Updates go through Sparkle, which the README says verifies an EdDSA signature and Developer ID. Because the code is public under an MIT license, a reader can check each of these statements.

What remains unverified

The project is young and small: the repository showed four stars and no forks when we read it, and we saw no independent security review. We did not build the app, inspect its network traffic or confirm that memory is cleared as described. The README also warns users to leave passwords and other secrets out of attachments, and notes the model cannot search the web, so answers come from training data.

Verdict: Signal for the pattern, not a privacy shield

Howler is Signal as a design pattern and Noise as a privacy promise. It suits throwaway questions on a shared or inspected Mac, especially paired with a local model. Do not use it with a hosted provider for anything sensitive, because local erasure does not reach provider logs. Anyone relying on it for more than convenience should read the repository, watch its network traffic, and check their provider's retention policy first.