Valutazioni e recensioni

What you hit was a training-data cutoff issue, not an app bug. Every LLM is trained on data up to a fixed point in time. After that cutoff, the models have no knowledge of newer releases or announcements unless live search is enabled. So if Opus 4.6 or 4.7 launched after their training window ended, the models could confidently conclude it was fake or nonexistent. The models you query also depend on which API keys and providers you have configured. AI Council is BYOK, so reasoning quality and knowledge recency vary between providers and model versions. AI Council includes built-in web search specifically for this case. In Settings → Web Search, add a Tavily or Brave API key and a globe button will appear in the composer. Enable it before sending your prompt and the app will retrieve live sources before deliberating. With web search enabled, the app would have found current references to those model releases and grounded the response in live information rather than historical data.

Thank you for updating your review - it genuinely means a lot to us. AI Council coordinates multiple providers, concurrent API calls, anonymised peer review across models, and file and image handling. Ensuring reliability across all combinations of models and settings is complex, and issues do occur as you found out. Your original feedback identified key gaps and directly influenced what was prioritised. Good to see the custom groups working - the speed versus deep-thinking split is a core use case. Thank you for your patience and please do feedback with any features you feel may be worthwhile adding (or indeed if you spot any more gremlins!).

Merci pour le retour honnête - nous le prenons au sérieux. La version 6.0, envoyée aujourd’hui en révision Setapp, corrige directement ce que vous décrivez : 1. Chaque erreur identifie désormais le fournisseur concerné, par exemple “Limite atteinte (anthropic)” au lieu d’un message générique. Plus d’échecs silencieux sans explication. 2. La synthèse du coordinateur diffuse les tokens en direct et relance automatiquement en cas d’erreur réseau transitoire. Auparavant, une session pouvait rester bloquée si un seul modèle bégayait - 6.0 récupère proprement. Pour les requêtes complexes, vous pouvez aussi ajouter votre propre clé OpenAI, Anthropic ou Google dans Réglages → Fournisseurs. AI Council l’utilisera avant de toucher les crédits Setapp, et vous isole ainsi sur un fournisseur fiable. Merci encore.

Sorry about that - this was a real bug. The default council required models from multiple providers, so entering a single API key left the council one model short of the minimum and the chat silently stayed disabled. Fixed in 5.3 (awaiting review): the council now auto-completes with starter models from whichever provider you configured, so a single key unlocks the app immediately. Please try again after the 5.3 update - thanks for flagging it.

感谢您的反馈！您提到的自定义 AI Provider 来源的功能其实已经在您评论之后的 5.2 版本中加入。 使用方法：打开「设置」→「Providers」标签页，滚动到列表底部，您会看到「Custom Providers」区域。点击右侧的 + 按钮，就可以添加任何 OpenAI 兼容的推理端点 - 输入名称、Base URL（例如 https://your-endpoint.com/v1）和 API Key 即可。这样您就能接入任何第三方推理平台。 即将发布的 5.3 版本会把 MiniMax 和 Z.ai（GLM）作为内置 providers 加入，中国用户不需要走自定义 provider 路径也可以一键配置。5.3 已经开发完毕，正在等待审核上架。 如果升级后在使用中还有问题，欢迎继续反馈。感谢您的耐心！ Thanks for the feedback. The custom AI provider feature you asked for was added in version 5.2, released shortly after your review. How to use it: Open Settings → Providers, scroll to the bottom of the provider list, and you’ll see a Custom Providers section. Tap the + button to add any OpenAI-compatible endpoint - enter a name, base URL (e.g. https://your-endpoint.com/v1), and your API key. You can route through any third-party inference platform this way. The upcoming 5.4 release additionally adds MiniMax and Z.ai (GLM) as first-class built-in providers, so users in China can configure them with one tap without needing the

Thank you for reporting this. You’re right that the app supports two local-model paths, and only one currently appears in the Council Members list. If you install Ollama (ollama.com) and run ollama pull <model>, any model you pull will appear automatically in the Council Members picker on macOS. This path works end-to-end for both direct chat and councils. The “Link Local Model” feature that scans your disk for .gguf files is a separate path intended to help users locate existing models, but it does not yet run inference against them. That is why they appear in the Local Models list but not in the Council Members picker. This is a gap on our side rather than intended behaviour. We will address it in the next update, either by making linked models fully usable or by making the limitation explicit in the UI so it is not misleading. Apologies for the frustration in the meantime.

The timeouts and dropped connections you are experiencing happen because AI Council runs multiple long-running API calls in sequence - gathering responses, ranking them, and synthesising a final answer. If a network interruption occurs anywhere in that chain, the entire step can fail. The next release will improve retry handling so transient connection drops trigger automatic retries with back-off instead of surfacing immediately as errors. If you are seeing this consistently, it is also worth checking your API provider’s status page. Some providers, particularly free tiers, have higher latency and lower reliability, which becomes more noticeable during multi-model runs. Connection handling is continuing to be hardened in upcoming releases.

Thanks for the feedback - and the Intel context. This review reflects v6, and a number of changes have shipped since. Two points are relevant to what you described: Between v6 and the current release, the per-request timeout was tightened to provide faster failure feedback. That addressed long spinner complaints, but it also made behaviour worse for slow reasoning models such as Claude Opus or o3-class thinking. In those cases, the provider has received the prompt but is still processing when the timeout is reached, which matches your symptom. The next release, currently in review, adds an automatic single retry on transient timeouts and dropped sockets for follow-up and continuation replies. It also normalises error messages so failures return a clear typed reason rather than a raw system string. Apologies for the frustration caused!

Gracias por la crítica honesta - la frustración es justa y lamentamos los créditos perdidos. Las alucinaciones son un problema real de todos los LLMs. En la versión 6.0, enviada hoy a revisión de Setapp, añadimos Búsqueda Web con 8 proveedores (Tavily, Brave, Exa, You.com, Kagi, SerpAPI, Perplexity, Google). Cada respuesta muestra las fuentes reales como enlaces clicables, para que puedas verificar de dónde viene la información. Para no gastar créditos de Setapp: en Ajustes → Proveedores puedes añadir tu propia clave de OpenAI, Anthropic o Google. AI Council la usará antes de tocar los créditos gestionados de Setapp. No existe una suscripción separada de AI Council. El cupo es parte de tu plan de Setapp AI, gestionado por Setapp. Gracias por la señal.

Thanks for the thumbs up, and for the honest feedback - you caught a real bug. A single transient 402 from the Setapp AI gateway was locking the “out of credits” state on for an hour, even after subsequent calls succeeded. So a momentary hiccup on MacPaw’s side would tell you you’d run out when you hadn’t. 5.3 (awaiting review) clears the exhausted flag the instant any request goes through successfully. Also worth knowing: Setapp AI credits are a separate pool from your main Setapp subscription (10/month on base, 125 on AI+ Enthusiast, 250 on Expert). 5.3 also uses cheaper models by default so credits stretch further. Please give it another try once 5.3 lands. Thanks!

Fair feedback - both were real. Both fixed in recent updates: Markdown tables: the table parser was rebuilt in 5.2. It now handles empty cells, inline formatting (bold, italic, code, links) inside cells, and column alignment syntax. UI jumping: two causes. The chat view re-parsed markdown on every update and used unstable list IDs, which caused flicker and scroll jumps - rewritten in 5.2 with native SwiftUI markdown and stable message identifiers. A conversation-switching bug in 5.3 (awaiting review) finishes the job on the remaining jumpiness around stage transitions. Please update and let us know how it feels. Thanks!

Thanks for flagging this - you are right, and the bug is on us. Mistral’s /v1/models endpoint reports its models with a type value our filter was not expecting, so the listing was being silently dropped to zero. Every Mistral chat model your key can access (mistral-large, medium, small, codestral, ministral, devstral, magistral, pixtral, open-mistral-nemo) was unreachable because of it. The fix is in the next release, now in review. Once it lands, deleting and re-entering your Mistral key, or hitting “Test” again, will pull the full catalogue. The same release also fixes the custom OpenAI-compatible provider issue and the “network connection was lost” pop-ups on follow-up replies. Sorry for the frustration caused.

Thank you for the honest review - the frustration is fair. Each Council question runs in stages and touches 3-4 models, so the base Setapp plan’s small credit pool can be consumed quickly. That is a real UX problem, not a trick. To be clear: there is no separate AI Council subscription. The credit allotment is part of your Setapp AI plan, set and billed by Setapp. We do not receive anything from an AI+ upgrade. To stretch your credits: add your own LLM provider key in Settings → Providers. AI Council routes through your key before ever touching Setapp credits, so a BYOK key on one provider cuts your credit burn per Council question significantly. You can mix BYOK providers with Setapp-routed ones in the same Council. 6.0, submitted for Setapp review today, also makes Apple Intelligence a first-class member on Apple silicon (free, on-device, no credits), clears the credit-exhausted banner when your balance recovers, and tightens several BYOK routing paths.

Thanks - all three bugs are fixed in 5.3 (awaiting review). Connection timeout: dropped per-request limits from 120 - 180s to 60 - 90s. Unreachable endpoints now fail fast so you keep flow. Conversation A showing B’s content: the orchestrator was holding stale results across conversations. It now resets on every switch, so A always shows A. “Generate Chair’s Report” unresponsive: errors from that retry path were silently swallowed. They now surface as a visible banner with the specific failure, and the cross-run state leak that caused the “one cycle then bugs out” pattern is gone. LM Studio: already supported. Start LM Studio’s local server, then in Settings → Providers pick LM Studio (Local). Your loaded models appear in the council picker, no API key needed. Thanks for the thorough report - it was genuinely useful.

Sticky “out of credits” banner: a single transient 402 from the Setapp AI gateway locked the banner on for an hour even after subsequent calls succeeded. Fix: the exhausted flag now clears as soon as any successful response comes through. Credit confusion: Setapp AI is a separate credit pool from your main Setapp subscription - base plans get 10 AI credits/month, AI+ Enthusiast 125, Expert 250. A single council question uses several calls. 5.3 adds clearer messaging and cheaper defaults so you do not burn through credits as quickly. Please try again once 5.3 lands. Thanks

Thank you for the review. A few clarifications: 1. “Setapp has no credits” - Setapp AI credits (MacPaw’s AI gateway) are tracked separately from your Setapp subscription credits. When the gateway returns 402, your AI allocation is exhausted even though your main Setapp account is active. Our error copy needs to make this clearer and we will update it. 2. “Keeps pointing to Setapp models with personal keys” - BYOK takes priority over Setapp AI automatically. Once you add an OpenAI or Anthropic key, those providers route through your key. The model IDs in the council look the same, but the routing path changes. A related bug affecting follow-up messages was fixed in 5.1. 3. “No local models for councils” - Ollama is supported on macOS. The council is pre-filled with Setapp AI models on first launch; open Settings > Council to swap them for your local models. We will make this more discoverable. A patch is in preparation to fix these points.