Most people using AI agents do not realise how much of their inbox, chat history and client files leaves their laptop every time they ask for a summary. For a lot of work that is fine. For a contract, a payroll sheet or a client's WhatsApp-style message thread, it is the reason you never let the agent near it.
On 20 October 2026, Perplexity gave Mac users a different option: Perplexity Computer can now run fully locally, with the model, the planning and the file processing all on your machine. The useful question is not "is local better?" It is "which tasks should stay on my Mac, and which should not?" Below is a simple three-lane routing rule, a copy-paste local-first prompt, and the honest limits.
What is Perplexity Computer's fully local mode?
Perplexity Computer's fully local mode runs the AI agent and its model entirely on an Apple silicon Mac. Task planning, reasoning and file processing happen on-device, and code or tool execution runs in isolated sandboxes. It launched on 20 October 2026 for Pro, Max and Enterprise subscribers, using Qwen 3.6 or PPLX Pro, a post-trained version of Qwen.
According to MacRumors' launch report, the local agent can work across native apps including Notes, Messages, Mail, Keynote and Numbers without sending personal data to the cloud.
Perplexity recommends 36GB of memory or more for local mode. Tbreak notes this is a recommendation rather than a hard minimum, but it is the figure to plan around, because the model, the agent and your files all share the same unified memory.
Local does not mean cut off. If a task needs current information, browser use or a cloud-connected app, you can authorise cloud access, and the local agent can call one of more than 15 frontier models for harder problems.
Setting it up takes three steps, according to Tbreak's summary of Perplexity's launch post:
- Update the Perplexity app for Mac to the latest version.
- Download a local model (Qwen 3.6 or PPLX Pro) inside the app.
- Choose the fully local option when you start a task.
Perplexity's own example is a good first task: review recent Messages and Mail threads, spot deadlines, and build an action list in Notes. It is private by nature, needs no web access, and is easy to check by hand.
How is local mode different from Hybrid Compute and the cloud?
The three modes differ in where the work starts and where your data goes. Cloud Computer runs on Perplexity's servers. Hybrid Compute, added on 1 September 2026, starts in the cloud and hands private data to local models. Fully local mode starts on your Mac and only reaches the cloud when you approve it.
- Cloud: strongest models and live web access, but your inputs are processed off-device.
- Hybrid Compute: frontier reasoning in the cloud, private files handled by local models. MacRumors reports it works on Macs with as little as 24GB of memory.
- Fully local: everything on the Mac by default. Recommended 36GB or more. Cloud is opt-in per task.
Perplexity also frames local mode as a way to control credit spend, because on-device steps do not draw on cloud usage. Tbreak's summary of Perplexity's post puts it simply: you pick per task, a private portfolio review on-device, or a deal review that needs live market research through Hybrid.
Which tasks should you keep on your Mac?
Keep a task local when its inputs include personal data, client confidential material or anything you would not paste into a public chatbot, and when it does not need live web information. Send it to Hybrid when it mixes private files with outside research. Use the cloud when every input is already public.
Here is the three-lane routing rule in practice:
- Lane 1, local: summarising your own Mail and Messages, extracting deadlines from client contracts, cleaning a staff roster in Numbers, drafting replies that quote customer details.
- Lane 2, Hybrid: comparing a client's internal sales sheet against current competitor pricing, or preparing a pitch from private notes plus fresh market news.
- Lane 3, cloud: researching a public topic, summarising published reports, browsing and comparing vendor websites.
For Hong Kong practitioners, Lane 1 maps neatly onto material covered by the Personal Data (Privacy) Ordinance: names, phone numbers, HKID details and employment records. Running locally reduces how far that data travels. It does not replace your own compliance checks.
Mixed tasks are where most people get this wrong. A monthly client report usually starts with private sales figures (Lane 1) and ends with market context (Lane 3). Split it into two runs: summarise the private numbers locally, then paste only the anonymised summary into a cloud task for the research.
A quick test before each task: if the input file were leaked, would someone outside your company be harmed or annoyed? If yes, start in Lane 1.
How do you write a local-first task brief?
A local-first task brief tells the agent exactly which files and apps it may touch, what to produce and where to save it, and what to do when a step needs the internet. Use five labelled lines: SCOPE, TASK, OUTPUT, CLOUD RULE and CHECK. The CLOUD RULE line is what keeps a private task private.
Try this prompt:
SCOPE: Only use Mail and Messages from the last 7 days, and the Numbers file "Client Tracker Q4" in my Documents folder. Do not open any other files or apps.
TASK: Find every message where a client asks for something with a date or deadline. For each one, record the client name, the request, the deadline and which app it came from.
OUTPUT: Create a new note in Notes called "Deadlines this week", sorted by deadline, earliest first. Add a final line listing any request where the deadline is unclear.
CLOUD RULE: Stay fully local. Do not search the web or call a cloud model. If any step needs the internet, stop and tell me which step and why, instead of going online.
CHECK: Before finishing, list how many messages you read, how many requests you found, and any message you skipped and why.
Two lines do most of the work. SCOPE limits what the agent can read, which matters because agents tend to explore. CLOUD RULE turns "stay private" from a hope into an instruction with a fallback.
The CHECK line gives you a quick audit trail. If the agent says it read 140 messages and found 3 requests, you know whether to trust the list or spot-check it.
The same structure works for browser agents. Our ChatGPT in Chrome guide uses a similar stop-line pattern for tab control.
What are the limits of running AI fully locally?
Fully local mode trades capability for privacy. The local models are open-weight Qwen variants, not the strongest frontier models, it needs a high-memory Apple silicon Mac, it still requires a paid Perplexity plan, and it only covers tasks that never need the internet. Expect to route harder reasoning to Hybrid or the cloud.
- Hardware. 36GB or more is recommended. Many work laptops ship with 16GB or 24GB, which points you to Hybrid Compute instead.
- Model strength. Qwen 3.6 and PPLX Pro run on your Mac. For complex analysis, Perplexity itself offers a route to more than 15 frontier models, which is a signal that local is not meant for everything.
- Subscription. Local mode is for Pro, Max and Enterprise subscribers, so you pay for a plan even when the model runs on your own hardware. Commenters on MacRumors raised exactly this point.
- Mac only. This release is for the Perplexity Mac app. Tbreak reports a separate Portable Computer for Windows RTX PCs.
- Region. Perplexity's launch post does not mention regional availability, so confirm it works on your account in Hong Kong.
Local also does not mean error-free. A local agent can still misread a date or miss a message, so keep the CHECK line and review anything it plans to send.
How can you test local mode in 20 minutes?
Run the same private task twice: once in fully local mode and once in Hybrid Compute, using the five-line brief above on one week of your own email. Compare accuracy, speed, and whether the local run ever asked to go online. That tells you which of your routine tasks can safely live in Lane 1.
- Did both runs find the same deadlines?
- Did the local run stop and ask before any online step?
- How long did each run take on your Mac?
- Did either run open a file outside SCOPE?
If local matches Hybrid on accuracy for your inbox triage, move that task to Lane 1 permanently. If it does not, keep the private steps local and let Hybrid do the reasoning. For the wider picture of how agents connect to your apps, see our guide to MCP and AI connectors.
The bigger shift is that "where does this task run?" is now a setting, not an assumption. Decide it per task, write it into the brief, and your private work stops being the work you never automate. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Details reflect Perplexity's 20 October 2026 announcement as reported by MacRumors and Tbreak. We have not benchmarked local mode ourselves. Plans, model options and requirements may change, so confirm in the Perplexity Mac app before relying on them.
Want AI that works on private data, safely?
Routing tasks by sensitivity is the first step. Running AI staff inside a secure, isolated environment for your whole team is the next. UD's AIStaff Solution deploys AI employees in a permission-controlled sandbox, and we'll walk you through every step, from tool setup to workflow design and deployment.