Perplexity Computer Runs Fully Local on Mac: Model and Agent Stay on Your Computer, Local Tasks Consume No Credits
From client file reviews and proposal budgets to cross-app to-dos, see what it can do for you, how fully local and Hybrid divide the work, and how privacy controls and requirements work.
Perplexity has launched a fully local mode for Perplexity Computer in its Mac app: both the model and the task-executing agent run directly on the Mac. With user authorization, it can process local files and native applications without sending sensitive content to cloud models for analysis.
Key highlights include:
- Both the model and agent run on your Mac: The local model understands tasks, while the agent calls approved tools and applications;
- Local processing of private materials: According to official documentation, local tasks can process sensitive information in files and apps without transmitting this content to cloud models;
- Local tasks do not consume credits: They do not consume Computer credits, though subscription requirements still apply;
- Optional Hybrid compute: The cloud handles public search and reasoning, while the Mac handles private files and local operations, with Computer automatically coordinating the two sides;
- Subscription and device requirements apply: Available to Pro, Max, and Enterprise users; requires a supported Apple silicon Mac, with 36 GB unified memory recommended for fully local execution.
See how client files turn into review notes, and how action items across email, messages, Slack, and calendar are written directly to Reminders. The video is an official feature demonstration.
Features: How Does a Local Agent Run Specific Tasks on Your Computer?
In day-to-day use, tasks assigned to AI assistants often involve substantial amounts of non-public personal information. Official demonstrations illustrate two typical local execution scenarios:
The first scenario is deep review of local sensitive files. The user selects the PPLX Pro model in the input box and prompts it to "read the latest emails and tax documents under ~/Clients/West/Okonkwo, prepare Q3 investment review notes for the Okonkwo family, and highlight quarterly changes along with core topics to cover for 2027." The screen shows spreadsheets, emails, and terminal tools before generating a review page detailing topics like retirement, housing, investment diversification, and Social Security. This illustrates the process of assembling discussion points from multiple sources, and cannot be used to assess the accuracy of financial advice.
The second scenario is collaborative extraction across system applications. The instruction directs the system to scan emails, text messages, Slack, and calendar to "add everything requiring a reply or action today to Reminders." The system automatically reads scattered information across apps—such as a text conversation about rescheduling a Friday meeting—and expands the to-do list from 1 item to 5 items, writing them directly into native Apple Reminders.
The official blog also describes a client proposal scenario: organizing confidential interview notes, client requirements, and an internal rate card into a proposal, generating a project scope, an Excel budget, and an Outlook email draft ready for human review.
The interface's "Where the work ran" clearly shows both model and agent ran on the Mac, with no cloud used. The resource numbers in the image reflect the web demonstration. Source: Perplexity Official Product Page, an official webpage snapshot, not a hands-on test for this article.
The common purpose across these tasks is bridging "reading material" directly into "preparing deliverables": saving users a round of manually extracting, organizing, and retyping across documents, spreadsheets, and applications. Anyone who routinely compiles client materials, builds budget proposals, or aggregates action items across apps can gauge whether it fits their workflow.
Local inference does not mean all applications function offline. Syncing or fetching messages in Mail or Slack may still require network access; this is distinct from where the AI interprets messages and determines the next step.
Core Mechanism: The Difference Between "Local Thinking" and "Local Doing"
"Local thinking" means the model understands tasks and analyzes content on the computer; "local doing" means the agent calls tools and manipulates files and applications on the computer. An action occurring on a Mac does not by itself prove that analysis also occurred on the Mac.
For instance, if a desktop assistant sends a tax return to a cloud model for analysis and then writes the results back to a local file, writing the file happens on the Mac, but analyzing the tax return still takes place remotely. Perplexity's fully local mode moves both model inference and agent execution onto the Mac.
- The local model handles understanding: Running directly on the Mac, the local model analyzes permitted materials and decides what to do;
- The local agent handles actions: It invokes approved tools, reads files, and processes application data to advance the task toward completion. Only when paired with this execution system does the model become more than answers in a chat box.
How Does Hybrid Connect Cloud Research with Private Data?
Hybrid lets the cloud handle public research, planning, and reasoning, while the Mac handles private files and native app actions. Computer automatically coordinates both sides and merges the work into a single deliverable. Users do not need to split tasks into multi-turn prompts or shuffle files across tools manually.
For example, when reviewing a transaction, the cloud can look up public financial disclosures, comparable deals, and market information, while the Mac reads confidential deal documents, extracts terms and financial assumptions, and flags discrepancies between both sides. That is the practical utility of Hybrid: bringing external research and internal materials that cannot be sent directly to the cloud into a single unified workflow.
The two subtasks divide work toward a single goal of updating the presentation; this is a snapshot from the official web component. Source: Perplexity Official Product Page, an official webpage snapshot, not a hands-on test for this article.
This process is governed by a Privacy Gate on the Mac that regulates outgoing information. According to official documentation, an on-device classifier first identifies sensitive values such as names, addresses, and account numbers, replacing them with aliases before transmission; when answers return to the Mac, real values are restored. For instance, a client's name is first replaced with "Client A," the cloud processes the request around the alias, and the returned output on the local machine swaps back the actual name. According to defined rules, the gate can also keep data local, refuse outbound transmission, or ask the user first. Hybrid still exchanges information with the cloud; its protection lies in governing what can leave and how it is transformed before leaving.
The Hybrid side shows local processing, cloud alias usage, and restoring real values upon return in sequence; the page also notes options to keep data local or request sharing permission. Source: Perplexity Official Product Page, an official webpage snapshot, not a hands-on test for this article.
How to Choose Between the Two Modes: Check Data Completeness and Whether Cloud Participation Is Needed
Fully local mode can handle complex, multi-step tasks; the choice does not hinge merely on how difficult a task is. Official documentation provides two clear comparisons:
- Private portfolio holdings plus already downloaded market data: suitable for fully local analysis. If all necessary materials are already on the local machine and you want planning, analysis, and execution to remain entirely on your Mac, choose fully local.
- Confidential financials plus fresh market research: suitable for Hybrid. The Mac processes confidential materials, the cloud supplies public research and reasoning, and the results are merged together.
When a local task needs fresh information, a browser, connected applications, or assistance from cloud models midway through execution, users can also authorize Computer to call the cloud; steps involving the cloud are then no longer processed fully locally.
First check what materials the task requires, then decide where to process them. Both modes handle multi-step workflows.
Both Materials and Analysis Stay on Your Mac
When to choose? All needed materials are already on the local machine, and you want planning, analysis, and execution to remain entirely on-device.
Example: Private holdings + downloaded market data → local analysis.
Local tasks consume no Computer credits, but still require a subscription.
Connecting Public Research and Private Materials into One Task
When to choose? You need to process confidential materials while also requiring fresh public research or cloud model participation.
Example: Confidential financials + latest market research → merged review.
Locally processed portions consume no cloud credits; this does not make the entire Hybrid task credit-free.
How Does Information Travel Back and Forth in Hybrid?
According to official documentation, the on-device classifier first identifies sensitive content like names, addresses, and account numbers before deciding how to handle it.
- ① Mac replaces sensitive valuesFor example, swapping a real client name for "Client A" before sending the request.
- ② Cloud processes the requestThe cloud uses the alias to perform approved research and reasoning.
- ③ Mac restores real valuesWhen the answer returns to the local machine, the original values are restored.
The Privacy Gate can also keep data local, refuse outbound transmission, or ask the user first based on rules; Hybrid still involves information exchange with the cloud.
When a fully local task requires cloud capabilities, it can invoke them with user authorization; steps involving the cloud are then no longer processed fully locally. Syncing messages across services like Mail or Slack may also access the network, which is distinct from where the AI conducts analysis.
Hybrid also enables remote usage: initiating a task from an iPhone, having the Mac read private files and execute sensitive steps, and returning results to the same task for access across devices. The product page also showcases recurring workflows executed every Monday, dispatching updated deliverables to Slack and email—suited for needs requiring continuously refreshed deliverables.
The results table separates local model rebuilding from cloud comparable deal research. Times shown are demonstration markers and should not be used to gauge actual runtimes. Source: Perplexity Official Product Page, an official webpage snapshot, not a hands-on test for this article.
What Is the Local PPLX Pro Model?
This demonstration used PPLX Pro, which is not the only local model option available. Perplexity describes it as a post-trained version of Qwen: continuing training on top of an existing base model for use in these tasks. The label "post-trained" alone does not disclose specific training methods or performance gains; Perplexity did not provide corresponding evaluation results in this announcement. Check the app for currently supported model options.
Who Can Use It, and How to Get Started?
Local tasks do not consume Computer credits, but the feature is available only to Pro, Max, and Enterprise subscribers. The former refers to whether executing local tasks deducts credits; the latter refers to access eligibility. The two should not be conflated. The Hybrid product page also notes that work handled by the local model does not consume cloud credits; this does not mean an entire Hybrid task is credit-free.
Hardware requirements vary by operating mode:
- Fully local mode: Runs on supported Apple silicon Macs. Perplexity recommends 36 GB unified memory; this is a recommended configuration, not a confirmed minimum memory requirement. Running models locally shares Mac resources with the operating system and other running applications; actual speed or compatibility on lower-memory machines cannot be inferred from this recommendation alone.
- Hybrid compute mode: Official product pages specify requirements of an Apple silicon Mac running macOS 15 or later, with a unified memory threshold of at least 24 GB and 32 GB recommended.
Getting started steps:
- Download and install the latest Perplexity Mac client from the official portal;
- Download the selected local model with one click following in-app prompts (configured once inside the app, with no need to install Ollama, set up local runtime environments, or configure API keys);
- In the model selector, choose fully local to run directly on the Mac, or choose Hybrid and assign the local model and cloud model respectively according to task needs. Check the app for current support details.
Members can copy or download the full Markdown, and let their AI subscribe and ask.
Found this useful? Leave your email — every new explainer, the moment it's out. Unsubscribe anytime.