On July 28, 2026, Perplexity brought Personal Computer to Windows — the same OS that powers roughly 1.4 billion machines. The new desktop agent reads local files, works inside File Explorer, Word, Excel, and Outlook, and routes each task to one of 20+ frontier models through a Model Council. For mobile-first teams like the ones we build for at Halmob, the interesting shift is not the desktop UI. It is that the same agent a knowledge worker starts on a phone can now finish the job on their PC without switching tools.
The launch was the lead beat in that day's smol.ai AINews newsletter, tracked alongside Moonshot's Kimi K3 release and Cursor's India Start on Grok 4.5. The theme threading through all three is the same: agents run asynchronously in the background, and the human supervises from whichever surface they happen to be near — phone, laptop, or Windows PC.
The 30-Second Version
What Personal Computer for Windows Actually Does
Personal Computer is Perplexity's agentic desktop. It launched on macOS in April 2026 and moved into Microsoft Teams and Microsoft 365 apps in May. The July 28 release extends the same agent to Windows, where it now has direct access to File Explorer, Outlook mail, Word documents, Excel workbooks, and any local folder the user grants. Perplexity gains the ability to combine that local context with Microsoft 365 data and public web results inside one conversational session.
Two properties matter for anyone building on top of it. First, the agent decides which model to call per step — a summary might go to a cheap open model, a code-heavy step to Claude or GPT, a reasoning step to Gemini. That routing is opaque to the user by design. Second, the agent runs long. A single "prepare the Q3 board review" prompt can span dozens of file reads, three or four API calls, and a final Word export — while the human checks in from a phone.
The Model Council: Multi-Model Routing Without the Wiring
The Model Council is Perplexity's name for the routing layer that picks a model per task. It is the same idea we described in the model routing layer for mobile AI agents in n8n flows, only shipped as a product feature instead of a piece of infrastructure you build yourself.
| Step in a session | Model that usually wins | Why the router picks it |
|---|---|---|
| Long-context file summary | Gemini or a long-context open model | Wide context window, cheap per token |
| Code or spreadsheet reasoning | GPT-5.x or Claude | Best pass-rate on structured logic |
| Search-grounded fact lookup | Perplexity's own retrieval stack | Native web index, citations included |
| Draft copy or email tone | Claude or a mid-tier open model | Natural voice, low cost per finished draft |
| Verification pass on the final output | A second model, different family | Cross-family checks catch single-model blind spots |
This is a real cost-shifting move. Under Perplexity's January 2026 three-year, $750M Azure GPU commitment, they can offer 20-plus model orchestration at a flat subscription price without every customer wiring their own routing table. The user pays $200/month and never sees the per-token bill; Perplexity absorbs the routing complexity in exchange for the subscription lock-in.
Why This Sits Squarely in the Mobile Supervision Pattern
The pattern the July 28 AINews issue called out — "work with agents from anywhere" — is now visible in three separate launches in the same week. ChatGPT Voice + Codex lets you supervise a coding agent by voice. Cursor Start (India, Grok 4.5) lets a mobile session hand off to a background coding agent. Personal Computer for Windows lets a desktop agent stay live while you check on it from the Perplexity iOS or Android app.
The desktop is where the agent works. The phone is where the human supervises. The two surfaces meet at the same session state.
We covered the same shape from a different angle in the Claude Cowork mobile, web, and background agents guide. Cowork optimizes for a coding workflow that starts on a laptop and continues on a phone. Personal Computer optimizes for an office-work workflow that starts on a phone and finishes on a PC. Same UX pattern, different job to be done.
How Windows Changes the Ceiling
Mac was a warm-up. Windows is the volume market — roughly 1.4 billion devices, most of them at desks that already run Word, Excel, and Outlook every day. The July 28 launch is the most direct challenge yet to Microsoft Copilot, on Microsoft's own OS, at six times the price. Perplexity is betting that a serious knowledge worker will pay more for a better router.
| Surface | Copilot for Microsoft 365 | Perplexity Personal Computer |
|---|---|---|
| Model choice | OpenAI models via Azure | 20+ models via Model Council |
| Local file access | Yes, inside Microsoft apps | Yes, File Explorer and Microsoft 365 |
| Web grounding | Bing with citations | Perplexity index with citations |
| Mobile supervision | Copilot mobile app | Perplexity iOS and Android |
| Starting price | $30/user/month | $200/user/month (Max plan) |
The price gap is the honest column. Copilot is the safe default for a company that already runs Microsoft 365 and does not care which model answers. Personal Computer is the premium tier for a team that wants model diversity, does its own vendor evaluation, and treats routing as a real skill.
What to Build When Your Users Land on Personal Computer
For a Halmob-style engagement — a mobile app, an n8n automation backbone, and an AI agent tier that spans the two — Personal Computer changes what the mobile app should own and what it should not.
- Do not rebuild the router. If your users already pay for Personal Computer at work, they are getting Model Council routing for free inside the desktop session. Your mobile app should hand off the task, not duplicate the routing.
- Own the trigger and the follow-up. The phone is the best surface for "start the job" (voice, photo, quick capture) and "confirm the outcome" (approve the final email, sign the doc). Leave the heavy execution to the desktop agent.
- Keep long-running work in n8n. Batch operations, downstream webhooks, and cross-partner sync belong in the automation layer. That keeps the desktop session focused on the live task the user is actually watching. We described the production shape of this split in the n8n on AWS ECS Fargate load test post.
- Instrument the handoff. If a session starts on iOS and finishes on the PC, log the join key on both sides so you can rebuild the timeline. The audit trail is what turns "agent did something" into "we know exactly what the agent did."
- Wire your own tools through Perplexity, not around it. If Personal Computer is the shell the user lives in, your mobile-first app should show up there as an integration, not as a competitor tab.
The Limits Worth Knowing Before You Deploy
- Price gates the audience. $200/user/month is a Max-plan barrier. Personal Computer is not going to reach the same seat count as Copilot in the first year. Design your integration for the top decile of users, not the median.
- Model routing is opaque. The Model Council does not tell the user which model handled which step. For regulated use cases — insurance, health, banking — you may need an audit path that logs the model per call, and Perplexity has not published one yet.
- Local file access is a governance question. Personal Computer reads folders the user grants it. That is a decision an IT department will want to shape before a rollout, not after.
- Data residency is unclear at the router layer. When the router picks a model, it also picks a hosting region. For EU customers, that has GDPR implications that are worth clarifying with Perplexity before a pilot.
- The mobile app is a supervisor, not a full desktop peer. Long sessions still need the PC to be awake for local file steps to run. Plan for the phone to be a check-in surface, not a full substitute.
Do Not Route Regulated Writes Through the Council Yet
Where This Slots Into the Halmob Stack
Most of what we build at Halmob sits at the meeting point of mobile apps, n8n automation, and AI agent orchestration. Personal Computer for Windows changes the middle tier of that stack for exactly one class of user: the knowledge worker who already spends the day in Word, Excel, and Outlook. For that user, the desktop agent takes over the heavy execution, the mobile app becomes the supervision and trigger surface, and n8n keeps handling the batch and cross-system work behind the scenes.
The pattern rhymes with the customer-journey completion agent we wrote about in the Ushur Agentic Platform post. Ushur completes an outbound customer journey; Personal Computer completes an inbound knowledge-work journey. Both leave the mobile app as the one place a human touches to start and confirm the work — and both push the heavy lifting off the phone.
When to Reach for Personal Computer and When Not To
A Simple Decision Rule
The Bottom Line
Perplexity Personal Computer on Windows is not a new model. It is a new default for how a knowledge worker talks to twenty models at once, from either their desk or their phone, without ever picking one. For mobile-first teams, that changes the job of the phone: it stops being the place where the agent lives, and becomes the place where the human keeps an eye on the agent that lives on the PC. Build the mobile app for the supervision job, keep the automation layer honest, and let the desktop agent do what it is now genuinely good at.
For source material, start with the SiliconANGLE report on the Windows launch, the Tech Times coverage of the 20-model routing, and the smol.ai AINews newsletter for the July 28, 2026 context on the wider "work with agents from anywhere" wave. To wire the mobile-supervision side of this pattern into a shipping product, Halmob pairs the mobile app, the n8n automation backbone, and the AI agent tier so the desktop agent, the phone, and the automation layer end up on the same session state.