AI Agents / March 2026
Perplexity Shipped a Software Agent
That Runs Your Mac 24/7.
Personal Computer is not hardware Perplexity manufactures. It is a $200/month software agent that runs on a Mac you already own, routing work across 20 AI models while you sleep.
On March 11, 2026, at its inaugural Ask 2026 developer conference in San Francisco’s North Beach, Perplexity announced Personal Computer: a persistent, locally anchored AI agent that runs 24/7 on a Mac you already own or purchase yourself. The product is software, not hardware. Perplexity does not manufacture any device. The agent runs on any Mac with macOS Sonoma or later; Perplexity specifically recommends a dedicated Mac mini because it runs silently and continuously without the power penalties of a laptop. A Mac mini M4 from Apple costs roughly $499 to $599. The Perplexity software itself requires a Max subscription at $200 per month. Perplexity opened waitlist access at the March conference and rolled out full access to all Max subscribers on April 16, 2026. Enterprise pricing runs $325 per seat per month, adding SOC 2 Type II compliance, SAML single sign-on, and integrations with over 400 business applications.
The Orchestration Architecture and Why It Matters
Layer 1: The host Mac (user-supplied hardware). Any Mac running macOS Sonoma or later works. Perplexity recommends a dedicated Mac mini because it operates silently, continuously, and at lower power draw than a laptop. The user buys the Mac from Apple at standard retail prices. Perplexity has no hardware product and charges no hardware fee.
Layer 2: Perplexity’s software agent. The agent installs on the host Mac and gains persistent access to the local filesystem, running applications, browser sessions, and system state. Pressing both Command keys activates it. The interface sits alongside standard macOS apps rather than replacing them. The agent operates in the background between activations, monitoring triggers and building context.
Layer 3: Model orchestration (cloud). Perplexity routes each task to the model it determines best suited across approximately 20 frontier models. Claude Opus 4.7 served as the default orchestrator as of April 2026. Users can also select GPT-5.4 or Claude Sonnet 4.6 as the primary orchestrator. The user does not choose model routing for individual subtasks. Perplexity does, dynamically.
How the Orchestration Model Works
Perplexity’s Personal Computer maintains persistent access to your local filesystem, running applications, browser sessions, and system state. Unlike cloud-based AI assistants that process individual requests statelessly, the Personal Computer agent maintains context across sessions. It knows what files you edited yesterday, what tabs you have open, and what applications are running.
The orchestration model routes queries across approximately 20 different frontier AI models, with no single provider exceeding 25% of total usage. This multi-model architecture reduces dependency on any single provider and allows task-specific routing: coding queries go to models optimized for code, research queries go to models optimized for reasoning, creative tasks go to models optimized for generation. The orchestration layer is Perplexity’s actual product. The models are interchangeable components. In an internal study Perplexity disclosed at the conference, the system handled the equivalent of 3.25 years of work in four weeks, with estimated labor savings of $1.6 million. These figures come from Perplexity’s own internal test rather than independent audit, and the methodology was not published.
The Business Model Problem
The business model follows the same subscription-lock pattern reshaping AI agent economics: a high-friction, high-commitment product (persistent agent with full filesystem access) that generates query data, API revenue from model partners, and a habit loop around Perplexity’s interface. At $200 per month, the economics are reversed from the subsidized-hardware model: Perplexity captures the recurring software margin while the user absorbs the hardware cost.
The comparison to Anthropic’s Cowork and Claude Code is direct. Cowork provides similar computer-use capabilities through a cloud-connected agent that does not require dedicated hardware. Claude Code provides persistent project context through a CLI tool that runs on your existing development machine. Both achieve overlapping functionality. What Personal Computer offers that cloud agents do not: truly persistent local context. Cowork connects when you invoke it. Personal Computer is always on, always monitoring, always building its understanding of your workflow. The question is whether that persistent awareness translates into enough additional value to justify $200 per month and the privacy implications of a continuously running AI agent with full system access.
The Privacy Equation
An agent with persistent access to your filesystem, browser history, application state, and running processes collects a detailed behavioral profile. Perplexity processes this data to improve its orchestration and personalization. The privacy policy governing what data leaves the device, what is processed locally, and what is sent to Perplexity’s servers or third-party model providers is the critical document that prospective users should read before installing the software.
The 20-model orchestration architecture means your data potentially flows to 20 different AI providers, each with their own data retention and training policies. Even if Perplexity does not train on your data, the query content sent to downstream model providers may be subject to those providers’ terms of service. Multi-model routing amplifies the privacy surface area: instead of trusting one provider, you are trusting twenty. Perplexity has not published detailed documentation on which data touches which providers.
What Is Not Yet Answered
Privacy architecture: All queries pass through Perplexity cloud. What data is retained, how long, for what purposes? Perplexity has not published a product-specific data policy as of April 2026. Offline capability: If Perplexity’s cloud is unavailable, what does the agent do? A product depending on continuous cloud connectivity carries reliability risk. Model transparency: Users will not know which model answers their query. When a model gives a wrong answer through Perplexity’s interface, who is responsible?
The competitive field for persistent AI agents (including memory-consolidation approaches like AutoDream) is crowded but unsettled. OpenAI’s Operator, Google’s Project Mariner, Anthropic’s Cowork, and Perplexity’s Personal Computer all target the same use case: an AI that can interact with your computer on your behalf. The differentiators are architectural (cloud vs. local), interactional (on-demand vs. persistent), and economic (subscription-only vs. subscription plus hardware). None have achieved sufficient reliability for unsupervised production use at scale. The winner will be determined not by which approach sounds best in a demo but by which one fails least often in the unpredictable chaos of real desktop environments.
Sources: Perplexity official Ask 2026 conference announcement, March 11, 2026; Perplexity blog; Digital Trends; Gadget Hacks; MacAppHub; CEO Aravind Srinivas public statements.
Updated 2026-08-18: Corrected product framing. Prior version incorrectly described Perplexity Personal Computer as dedicated hardware Perplexity manufactures at approximately $299. The product is software. Perplexity manufactures no hardware. The agent runs on a user-supplied Mac (any Mac with macOS Sonoma or later; Mac mini recommended). Subscription price is $200/month. Hardware cost (Mac mini) is borne by the user at standard Apple retail prices (~$499-599). Layer architecture updated accordingly.