Perplexity opened its desktop AI agent to a much wider audience this week. Perplexity Personal Computer for Mac, previously gated to the $200-per-month Max tier, is now available to every Pro subscriber and rolling into Enterprise as of May 7, 2026. The agent operates inside a native Mac app, edits local files, drives Mac apps, runs voice and dictation, and orchestrates work across more than 400 connectors. It is the most ambitious mass-market desktop AI agent shipped to date and a direct shot across the bow of OpenAI’s Operator and Anthropic‘s Computer Use.
What’s actually new
The product itself is not brand new — Personal Computer launched on the Max tier in April — but the May 7 release does three things that change the picture. First, it makes Personal Computer available to all Pro subscribers ($20 per month) and Enterprise customers, expanding the addressable user base by roughly an order of magnitude. Second, it ships inside a fully redesigned native Mac app with new orchestration, dictation, and multi-window UX. Third, it adds new file-system primitives that let the agent operate on your actual local files through a sandboxed bridge.
The architectural pattern is the part worth understanding. Perplexity Personal Computer runs the agent’s reasoning loop in the cloud (on Perplexity’s servers) but executes filesystem and application actions on the user’s Mac through a local bridge. The bridge is a small process the Mac app installs; it exposes the file system and a curated set of Mac app capabilities to the cloud agent under a permissions model the user controls. Every action the agent takes is logged, sandboxed, and reversible. Files modified by the agent live in a secure scratch area until you explicitly accept the changes.
The capability surface is broad. The agent can read and write files, navigate the web, take screenshots and reason about them, dictate from microphone input, queue and resume multi-step workflows, and drive applications through accessibility APIs. The 400+ connectors include the major SaaS surface: Gmail, Google Workspace, Notion, Linear, Jira, Slack, Asana, Salesforce, HubSpot, GitHub, Figma, Airtable, Stripe, and dozens of others. Connectors expose typed actions to the agent, not just web automation; the agent reads a customer record from Salesforce as structured data, not by clicking buttons.
The model behind the agent is Perplexity’s orchestration layer running on GPT-5.5 by default, with the ability to switch in Claude Opus 4.7, Gemini 3.5 Pro, or a handful of open-weights models for specific workloads. Perplexity has been explicit that the platform is model-agnostic; the orchestration is the differentiator. The agent picks the right model per turn based on task complexity, with the user able to lock a model if they prefer determinism.
One detail worth flagging: the Personal Computer agent has memory that persists across sessions, scoped to the user. The agent remembers your preferences, your file naming conventions, the way you like email replies drafted, and the people you correspond with most often. The memory is editable, exportable, and deletable from the app’s settings.
Why it matters
- Desktop AI agents just went mainstream. Personal Computer was a Max-tier product for a small audience; opening it to Pro brings desktop AI agency to roughly a million users overnight, with practical workflows that previously required custom builds.
- The local file model is a category shift. Cloud chat agents are useful; agents that read and write your actual local files are a different category. Anthropic’s Computer Use and OpenAI‘s Operator have both been moving toward this; Perplexity’s Mac-native bridge is the cleanest implementation shipping today.
- Voice plus filesystem plus connectors is a real productivity stack. Dictating a task and watching the agent execute it across local files, web tabs, and SaaS connectors closes a loop that has been promised since the original Siri demo in 2011. It is finally working.
- The competitive pressure on Apple is immediate. Apple Intelligence is still rolling out third-party model support in iOS 27. Perplexity is shipping the agent experience Apple has been promising for years, on Apple’s own hardware, today.
- Pricing is the unlock. At $20 per month for Pro, Personal Computer brings desktop AI agency into the same price bracket as ChatGPT Plus and Claude Pro. The decision is no longer “is this worth $200 a month” but “is this worth the same $20 I am already considering paying.”
- The auditable-and-reversible action model raises the bar. Other desktop agents have struggled with trust. Perplexity’s sandbox-plus-explicit-accept pattern is the right answer for getting users to actually let an agent touch their files.
How to use it today
Adoption takes ten minutes. The most useful early-day workflows are the ones that span local files, the web, and one or two SaaS connectors at once, because those are the workflows where the agent saves real time. The path below is the one we recommend for a first session.
- Subscribe to Perplexity Pro if you do not have it. Pro is $20 per month; Enterprise is custom; Max remains at $200 per month and adds higher rate limits.
- Download the Perplexity Mac app from perplexity.ai/mac. The new app is a separate download from the old menu bar utility; install both if you used the prior one.
- Sign in with the account that holds your Pro subscription.
- Grant filesystem permission on the prompts the app surfaces. The first prompt scopes which folders the agent can access; start narrow (Documents, Downloads, a project folder) and expand as you build trust.
- Connect one or two connectors. Gmail and Google Drive (or their Microsoft equivalents) cover most knowledge-worker workflows. Add Linear or Jira if you live in product. Add Salesforce or HubSpot if you live in sales.
- Run your first multi-step task. Try something concrete: “Read the latest three customer feedback documents in my Customer Research folder, summarize the recurring themes, and draft a post for the team Slack channel.”
- Review the plan before letting the agent execute. Personal Computer surfaces a structured plan before running. Accept, edit, or reject; the agent only acts after you approve.
- Inspect the actions. Every action is logged in the activity timeline. Reversible actions show an explicit undo. Files modified live in a scratch space until you accept.
For developers who want to script against the same orchestration layer via Perplexity’s API, the agent endpoint exposes a similar surface programmatically. The minimum viable call below kicks off a workflow that operates on a local file and produces a structured output.
import requests, os
PERPLEXITY_KEY = os.environ["PERPLEXITY_API_KEY"]
resp = requests.post(
"https://api.perplexity.ai/v1/personal-computer/sessions",
headers={"Authorization": f"Bearer {PERPLEXITY_KEY}"},
json={
"instruction": (
"Read every .md file in ~/Documents/customer-research, "
"extract recurring product complaints, and produce a "
"ranked list with quoted examples. Save as research-themes.md."
),
"model": "auto",
"filesystem_scope": ["~/Documents/customer-research"],
"connectors": [],
"approval_mode": "plan_only",
},
timeout=60,
)
session = resp.json()
print("Session:", session["id"], "Plan:", session["plan"])
The approval flow lives over the same API. After reviewing the plan, you confirm to start execution.
requests.post(
f"https://api.perplexity.ai/v1/personal-computer/sessions/{session['id']}/approve",
headers={"Authorization": f"Bearer {PERPLEXITY_KEY}"},
)
How it compares
The desktop AI agent category sorted itself into a clean set of options over the last six months. The table below compares the major contenders as of mid-May 2026.
| Product | Where it runs | Underlying models | Price | Filesystem access | Best at |
|---|---|---|---|---|---|
| Perplexity Personal Computer | Native Mac app + cloud orchestration | Multi-model (GPT-5.5, Claude, Gemini) | $20/mo Pro or $200/mo Max | Yes, sandboxed and reversible | Multi-step workflows across local files, web, connectors |
| OpenAI Operator | Cloud-only browser agent | GPT-5.5 / GPT-5 | Included in ChatGPT Pro ($200/mo) | No (web-only) | Web automation tasks |
| Anthropic Computer Use | API for developers; experimental UI | Claude Opus 4.7 / Sonnet 4.6 | API token usage | Yes (developer-managed) | Custom desktop automation builds |
| Apple Intelligence (iOS 27) | On-device + Private Cloud Compute | Apple-foundation + selectable third-party | Free with hardware | Limited to Apple-native data | Privacy-first iPhone and Mac workflows |
| Microsoft Copilot Studio | Windows + cloud | GPT-5 family via Azure | $30/mo per user (Copilot) | Yes, on Windows | Microsoft 365-anchored workflows |
| Rabbit R1 Teach Mode | Dedicated device + cloud | Proprietary multi-model | Device purchase + subscription | No (browser-emulated) | Niche automation experiments |
The decision tree is simple in 2026. For Mac users who want broad desktop agency at a normal price, Perplexity Personal Computer is the new default. For ChatGPT Pro users running browser-only automation, Operator is still the obvious answer. For Windows-anchored organizations, Microsoft Copilot remains the right starting point. For developers building custom desktop automation, Anthropic’s Computer Use API is the right substrate. Apple Intelligence is the privacy-first option but not yet a true competitor to the others on capability breadth. The category is no longer about whether desktop AI agency exists; it is about which agent fits your platform and your budget.
What’s next
Three threads to watch between now and Apple’s June developer conference. First, expect Apple to respond at WWDC 2026 with a deeper third-party agent integration story for iOS 27 and macOS 27, since Perplexity has now seized the Mac agent narrative Apple was likely planning to claim. Second, expect a Windows version of Perplexity Personal Computer within ninety days; Perplexity has signaled the work is well underway and the Mac launch is the wedge. Third, expect a major OS-level integration push from the model labs — Anthropic’s Computer Use moving toward a managed UX, OpenAI extending Operator beyond the browser, Google bringing Gemini’s agent to ChromeOS and Android with first-class desktop semantics.
The longer arc is that the AI assistant is moving from a chat window to an operating system layer. Personal Computer is one of the first credible products to actually live there. The user no longer opens an AI tab; the AI is the surface across which the user does work. The next eighteen months will sort out who owns that surface across the major operating systems, and the ownership economics are genuinely large.
Frequently Asked Questions
Does Personal Computer work on Windows or Linux?
Not yet. The May 7 release is Mac-only. Perplexity has signaled a Windows version is in development; no date has been confirmed. There is no public Linux roadmap.
Do I need the Max plan, or is Pro enough?
Pro ($20 per month) is sufficient for full Personal Computer access. The Max plan ($200 per month) adds higher rate limits, priority access to the heaviest models, and longer context windows for very large workflows. Most users will not feel the Max constraints; engineers and power users running long-running agentic workflows may.
How does Perplexity handle privacy of my local files?
The agent’s reasoning runs on Perplexity’s servers; file contents the agent reads or writes are transmitted to those servers under the standard Perplexity data handling terms (no training on customer data on paid plans, plan-level data residency options). The local bridge sandboxes filesystem access to folders you explicitly approve. Personal Computer is not appropriate for sensitive data you cannot send to a third-party cloud; for those workloads, use a fully on-device tool.
Can Personal Computer be reversed if it makes a mistake?
Yes. Every action is logged with explicit undo where possible. File modifications happen in a sandboxed scratch area until you accept; rejected changes are discarded. Connectors with destructive actions (sending email, posting to Slack, modifying CRM records) prompt for explicit confirmation before execution unless you have configured otherwise.
What is the difference between Personal Computer and the regular Perplexity app?
The regular Perplexity app answers questions and runs web searches. Personal Computer adds the agent’s ability to act on your behalf: read and write local files, drive native Mac apps, run multi-step workflows across web and connectors. The two surfaces are bundled in the new Mac app; you choose which mode to use per task.
Will this replace tools like Raycast or Alfred?
Personal Computer overlaps with launcher-style tools on some tasks (quick file search, app launching, dictation) but is fundamentally different. Launchers are deterministic shortcut systems; Personal Computer is an agent that plans and executes. Many users will run both, using the launcher for predictable single-step work and Personal Computer for ambiguous multi-step work.