Cai vs Perplexity Assistant for Android

“I set up Cai with a custom action to take a stack trace from my clipboard and open a pre-filled GitHub issue in 10 minutes. The Ollama backend means I can use a larger local model when I'm at my desk and fall back to Ministral 3B on the go. MIT license means I can fork it and add my team's internal tools.”

“Ministral 3B is fine for basic text tasks but it stumbles on anything requiring real reasoning or domain knowledge. Most users will hit its limits quickly and need to set up Ollama anyway — which is a non-trivial setup process for non-developers. The privacy story is genuine but the capability bar is lower than what cloud alternatives offer.”

“The category is proactive mobile assistant, and the direct competitor is Google Assistant — which Google has been slowly cannibalizing with Gemini while leaving a genuine gap in reliable on-device action-taking. Perplexity's bet is specific: they're wagering that their search quality and model integration is good enough to own the default assistant slot on Android before Google locks it down with Gemini natively. Where this breaks is power users with complex multi-app workflows — the moment you need it to draft a reply, attach a file from Drive, and schedule a follow-up in one shot, current on-device agent reliability falls apart. What kills this in 12 months isn't a competitor — it's Google shipping Gemini as a mandatory default assistant in Android 16 and closing the third-party assistant API surface. To be wrong about that, Google would have to lose an antitrust battle specifically over assistant defaults.”

“Cai represents a class of tools that become dramatically more useful as on-device models improve. When Bonsai-scale 1-bit models hit 8B+ quality at 131 tokens/sec locally, Cai's architecture is exactly right — a minimal, composable action layer on top of local inference. The MIT license means the community will build the plugin ecosystem.”

“The thesis here is falsifiable: within 36 months, the OS-level assistant slot becomes the most valuable piece of real estate on mobile, and whoever owns it owns the user's intent graph. Perplexity is betting that the assistant layer decouples from the OS manufacturer before Google can re-couple it with Gemini — a real race with a real dependency on regulatory pressure and Android's openness persisting. The second-order effect nobody is talking about: if Perplexity's assistant accumulates enough behavioral data from proactive actions — calendar patterns, app usage, translation contexts — they build a personalization moat that their search product has never had. The trend line is the shift from reactive query-response to ambient intent capture; Perplexity is on-time, not early, but they're one of the only non-platform players with the model quality to make it credible.”

“I've been looking for a way to do quick AI rewrites and tone adjustments in any app — not just in a web browser — without pasting things into a chat interface. Cai works in Figma, Notion, Miro, everything. The local privacy angle matters a lot when I'm working on client content that's under NDA.”

“The buyer here is the consumer who decides to swap their default assistant — a notoriously hard conversion that historically requires either zero friction or a viral forcing function, and this has neither. The pricing architecture is a problem: free tier commoditizes the product against Google's free default, and $20/mo Pro is a hard sell when the incumbent costs nothing and is already on the device. The moat question is the real issue — Perplexity's defensibility in search was always distribution, not model quality, and on Android they're fighting for distribution against the platform owner. When Google ships proactive Gemini actions as a system-level feature in a quarterly Android update, Perplexity's action layer becomes a third-party workaround. What would need to change: a carrier or OEM distribution deal that makes Perplexity the default out of the box, which is exactly the kind of deal Google's agreements with OEMs historically prevent.”

“The job-to-be-done is clear and single-threaded: be the assistant that both answers and acts without making you switch apps. That's a real job, and current Google Assistant does it poorly enough that there's genuine hire-me potential here. The onboarding concern is real — setting a third-party app as the default assistant on Android requires navigating Settings sub-menus that most users abandon before completing, which means Perplexity has to earn the switch before they can deliver value, a sequence that's backwards from good onboarding. The product opinion is there: Perplexity has bet on proactive and ambient over reactive and query-based, which is a genuine point of view. The gap between what's shipped and what's needed is reliable multi-step action completion — one failed calendar creation or misread translation and users revert to the default, and that trust window is narrow.”