Phi is a macOS-only AI browser developed by Phinomenon. It implements the Chromium engine as a native macOS app built on AppKit/SwiftUI, keeping the personal memory it learns from your browsing on your device while letting you delegate work to agents in progressive stages according to how much you trust them. Released in April 2026, it is a new browser that covers everything from memory management to agent automation in one package.
Key Features
- Local-first Memory: Phi automatically learns context from your browsing behavior and stores it as files on your device, where you can view, edit, delete, and export it. It is not sent to the cloud and is not used for model training (per the official FAQ).
- A named AI assistant: The assistant you name during onboarding answers questions and produces summaries using the context of selected text, tabs, and tab groups.
- Progressive agent automation: You can delegate work to AI at three levels — on-demand actions you watch happen, Shadow Tasks that run in the background after approval, and scheduled execution through the resident app “Phi Sentinel.”
- Skills system: Teach a workflow once in the open SKILL.md format, and you can reuse it afterward with a single sentence.
- Spaces, Profiles, and URL Rules: Sidebar-centric workspaces (Spaces) work alongside Profiles that separate cookies and logins, plus URL Rules that automatically route pages based on domain conditions.
- Phi Link (Telegram integration): Chat with your assistant while away from your desk and receive task result notifications through the official bot or your own Telegram bot.
- Open-source core and external control: The core is published under Apache 2.0, and external clients such as coding agents can drive the browser through a Playwright-compatible CLI or MCP.
Pricing
| Plan | Monthly price | Key features |
|---|---|---|
| Free | Free | All AI features available at no cost, including Memory, the assistant, agent actions, and Skills. Phi Cloud models are the default, but you can optionally switch to your own API keys or local models (Ollama, LM Studio, etc.) |
Pricing reflects information as of September 2026. Free is currently the only plan offered. The FAQ hints at possible paid tiers for advanced features in the future, but neither prices nor the features involved have been announced. Check the official site for the latest information.
Pros and Cons
✅ Pros
- Memory learned from browsing behavior stays only on your device rather than in the cloud, and is not used for model training (stated explicitly in the official FAQ)
- Agent autonomy is split into three levels — immediate execution, approval-gated background tasks, and scheduled execution — so higher-risk actions can be confirmed each time
- The Chromium engine is implemented as a native macOS app using AppKit/SwiftUI (described as lighter than Electron-based wrappers, though this is Phi’s own claim with no third-party measurements to back it up)
- The core is open source under Apache 2.0, and it can be driven from coding agents and similar clients via CLI or MCP
⚠️ Cons
- macOS only (Apple Silicon and macOS 14 or later required); Windows and Linux are not supported
- Free for now, but the official FAQ hints at paid advanced features in the future, so the pricing model is not settled
- Released in April 2026, so its extension ecosystem and track record are still thin compared with major browsers like Chrome
- It bills itself as “local-first,” but heavier AI processing still relies in part on frontier models in the cloud
Comparison with Similar Services
The AI browser category includes competitors such as Dia (The Browser Company’s successor to Arc), Perplexity Comet, Arc itself (whose development has slowed), and the philosophically similar privacy-focused Aside.
That said, information about how Phi differs from each of these comes mainly from comparison posts on Phi’s official blog. Since those are first-party comparisons with no neutral third-party evaluation to confirm them, this article stops at noting that competitors exist rather than tabulating who wins on which point. When you actually evaluate these, trying each one against your own use case is the better approach.
Who It’s For
- People looking for an AI browsing environment that keeps the behavioral data accumulated from personal browsing history inside their own Mac rather than in a vendor’s cloud
- People who want to run recurring automation tasks — price monitoring, page change detection — that should continue even with the browser closed
- Developers who want to hand browser control to coding agents through a Playwright-compatible CLI or MCP
- Former Arc users who, now that Arc has entered maintenance mode, are looking for a destination that carries over Space- and Profile-based workspace management
Summary
The deciding question is whether you are comfortable handing the behavioral data your browser accumulates to a vendor. If you are, you have plenty of other options; if you want that data to stay on your machine while still using agents, the field narrows considerably. On the other hand, OS support is limited to macOS, the browser is only months old, and with a free plan as the only tier, its future direction remains undecided.