Perplexity Computer: Complete Guide to the AI Agent, Credits and Workflows

by David Porter

Friday, 31 July 2026 at 15:51

thumbnail_perplexity-computer-ai-agent-c

Perplexity Computer is an AI agent that completes multi-step work across the web, files and connected applications. Unlike the standard Perplexity experience, which primarily finds and synthesizes information, Computer can build an output, use a tool, update a system, send a message or continue working on a schedule.

Perplexity expresses the difference in six words:

Perplexity answers. Computer does your work.

That is an aspiration as well as a useful product boundary. Computer can coordinate models, subagents and tools, but it does not acquire human accountability. Its work can be incomplete, expensive, inaccurate or inappropriate for a connected system. A successful deployment requires bounded tasks, least-privilege connectors, checkpoints and cost controls.

This guide explains the cloud product, Computer credits, scheduled tasks, reusable Skills, Brain, enterprise administration and the newer Personal Computer for Mac. For Perplexity’s search and research tools, use the complete Perplexity guide.

Perplexity Computer at a glance

Product type	Cloud AI worker with models, tools, memory, connectors and schedules
Availability	Active paid subscribers; web, iOS and Android
Primary input	A natural-language objective plus files, web and approved connected context
Outputs	Research, documents, slides, spreadsheets, apps, messages, updates and other task results
Execution	Background and attended cloud tasks in isolated environments
Connectors	Hundreds of supported services, subject to plan and availability
Recurring work	Scheduled Tasks
Reusable procedures	Skills
High-end memory preview	Brain for eligible Max users
Local desktop version	Personal Computer for macOS 14 or later
Billing unit	Computer credits; 100 credits equal $1
Primary risk	A capable agent can take the wrong action across real systems

What Computer can do

Computer combines several capability classes.

Research

It can search widely, divide a topic among parallel subagents, inspect current sources and produce a synthesized result.

Create

It can produce reports, presentations, spreadsheets, visual material and functioning applications. The quality and format depend on the task and available Skills.

Use tools

It can operate a browser, execute code in a sandbox and call supported tools. Tool use turns a language response into an action sequence.

Use connected applications

Published examples include Gmail, Outlook, Slack, Notion, GitHub, Linear, Salesforce, Snowflake and Databricks. The available connector catalog is larger and changes over time.

Remember context

Persistent memory can retain preferences and relevant work context. Brain, a research preview for eligible Max users, is intended to build a deeper working model of projects, people and files.

Run in the background

Tasks can continue asynchronously. Scheduled work can monitor conditions, create briefings and notify the user when something relevant happens.

How Computer works

A complex instruction is typically decomposed into stages:

interpret the desired outcome;
plan subtasks;
select models, Skills and tools;
obtain required context;
execute steps, sometimes in parallel;
combine results;
create or update the output;
request attention when a blocker or consequential choice appears.

Perplexity describes specialized subagents for research, finance, news monitoring and coordination. The user does not need to choose every underlying model manually.

This architecture is valuable when work is compositional. A single assignment can involve research, analysis, slide creation and an email draft.

It is also why instructions need boundaries. One vague objective can create many model calls, searches, retries and tool actions.

Computer versus Search, Research and Create

Product	Primary job	Typical stopping point
Search	Answer a question with sources	Cited response
Research	Investigate a substantial topic	Research report
Create	Produce a file or asset	Finished deliverable
Computer	Coordinate information, tools and actions	Outcome across one or more systems

Use Research when the work ends with understanding. Use Create when it ends with an asset. Use Computer when the workflow requires orchestration or an external action.

Do not involve Computer in a task merely because it is newer. Search is faster and cheaper for a fact.

How to access Computer

Perplexity says Computer is available to users with an active subscription through the web, iOS and Android.

In the main Perplexity interface:

select Computer;
choose Start a task;
describe the objective;
attach or connect only the needed context;
inspect the proposed plan and permissions;
run the task;
review the result and credit use.

The connector panel lets users enable, install and disconnect supported services.

How Computer credits work

Computer uses credits rather than the ordinary search allowance. Standard Ask searches do not consume them.

Perplexity defines:

100 credits = $1

Cost varies with task complexity. More models, tools, media generation, searches, retries and actions can consume more credits.

At review, the plan-level starting points included:

Pro: no standard recurring monthly allocation, with a possible one-time 4,000-credit bonus;
Max: 10,000 monthly credits, with a possible one-time 35,000-credit bonus;
Enterprise Pro: 500 monthly credits;
Enterprise Max: 15,000 monthly credits.

Introductory credits and exact enterprise bonuses can vary. The account credit screen is authoritative.

What happens when credits run out?

Perplexity says active Computer tasks pause rather than cancel. They can resume when credits become available. A scheduled run can be skipped and retried before the normal schedule resumes.

Auto-refill

Auto-refill can purchase more credits when the balance falls below a threshold. Leave it off during the pilot, or set a conservative spending cap and alert.

Estimate before automating

Perplexity’s published examples place light tasks in a rough 100–350 credit range and very large work in a broad 2,400–9,800 range. These are not quotes.

Run a task three times with stable instructions. Record median cost, maximum cost and failure cost. Then estimate the recurring budget.

The Perplexity pricing guide includes the complete plan and credit tables.

Write a good Computer task

Use an outcome contract:

Outcome: [the final state or deliverable] Inputs: [approved sources, folders, records or accounts] Actions allowed: [read, draft, create, update] Actions requiring approval: [send, publish, delete, spend, change access] Quality checks: [tests, citations, reconciliation] Budget: [credit ceiling or stop condition] Finish: [where to save and what completion means]

Example

Outcome: Produce a five-slide weekly competitor brief. Inputs: The named competitor list, public web sources and the approved pricing spreadsheet. Actions allowed: Search, read the spreadsheet and create a PPTX. Approval required: Do not email, publish or update the spreadsheet. Quality checks: Use primary sources for product and price changes; show source dates; exclude repeated news. Budget: Stop and ask before exceeding 900 credits. Finish: Save the deck and a source log in this session.

That prompt is easier to supervise than “handle our competitor intelligence.”

Twelve practical Computer workflows

1. Weekly market brief

Research named companies, identify only material changes, create a concise brief and save a source log. Keep distribution manual until quality is stable.

2. Sales-account preparation

Read approved CRM fields and public company information, then prepare a meeting brief. Do not alter opportunity stages or contact the prospect.

3. Inbox triage

Classify selected messages, identify deadlines and draft replies. Exclude sensitive folders and require approval before sending.

4. Project status report

Combine approved Linear, GitHub or project records into progress, blockers and next actions. Link every status claim to the underlying item.

5. Research-to-presentation

Investigate a bounded question, produce an evidence matrix, create slides and retain notes. Separate factual approval from visual approval.

6. Dataset enrichment

Add public attributes to a defined set of rows. Write source and date beside each value. Never overwrite the original dataset.

7. Customer-feedback synthesis

Read an approved export, remove direct identifiers where possible, cluster themes and quote representative language. Do not infer demographic attributes.

8. Release-note monitor

Check selected vendor documentation on a schedule and notify only when a product change affects named dependencies.

9. Vendor questionnaire preparation

Gather public security documents and prefill a review worksheet. Mark contract-dependent answers as unresolved rather than guessing.

10. Content evidence pack

Build a verified source bundle and outline for an article. Keep original writing, legal review and publication under human control.

11. Small internal application

Create a prototype that transforms approved input into a report. Test with synthetic data before connecting production systems.

12. Travel or event coordination

Research options and draft an itinerary. Require approval before booking, paying, cancelling or disclosing identity details.

Scheduled Tasks

Scheduled Tasks let Computer run recurring work in the cloud even when the user’s laptop is closed.

Supported schedules can be hourly, daily, weekdays, weekly, monthly or another recurring pattern, but not more frequent than hourly.

Typical uses include:

price or policy monitoring;
inbox or pipeline checks;
daily and weekly briefings;
recurring data collection;
news or competitor alerts.

Background versus attended tasks

A background run starts in a fresh isolated agent and may not inherit the full creating conversation. It is suitable for monitoring and reports.

An attended task can be needed when work relies on richer conversation context, drives a browser, creates a complex file or updates a hosted resource.

Put every important baseline in the scheduled instruction or an approved connected source. Do not assume a fresh run remembers an unstated decision.

Scheduled-task cost control

Every run uses credits. An hourly task can execute roughly 24 times as often as a daily one.

Use:

Run each weekday at 08:00. Check [sources] for changes since the previous successful run. Notify only if [material condition] occurs. If there is no material change, produce no alert. Do not use more than [budget] credits per run.

At review, each Computer conversation could own up to 15 scheduled tasks. There was no pause state; stopping a task removed it and it had to be recreated later.

Computer Skills

Skills are reusable instruction sets that teach Computer how to perform a type of work. Built-in examples cover research, reports, slides and other outputs. Computer can combine several Skills in one workflow.

Users can create a custom Skill conversationally or upload one in a supported form.

What belongs in a Skill

scope and trigger;
required inputs;
approved source classes;
procedure;
output schema;
validation;
prohibited actions;
escalation conditions.

What should not be hidden in a Skill

credentials;
a broad instruction to bypass approval;
sensitive data not needed by the task;
rules that contradict organizational policy;
an unreviewed action on every matching query.

Treat Skills like operational procedures: version, test and review them.

Brain

Brain is a research-preview memory system for eligible Max subscribers using Computer. Perplexity says it builds a working model of projects, people and files so the agent can act with more context.

Persistent context can reduce repetitive setup. It can also preserve outdated assumptions or sensitive associations.

Users should be able to answer:

What did Brain remember?
Which source did the memory come from?
Can it be corrected or deleted?
Which task used it?
Does it remain appropriate after a project or role changes?

Do not use remembered context as the sole authority for a current instruction or access decision.

What is Personal Computer?

Personal Computer is Perplexity’s rebuilt Mac application centered on Computer.

It adds:

access to user-approved local folders;
reading, writing and searching local files;
control of supported native apps such as Mail, Finder, Slack, Messages and Notes;
a floating task bar;
voice control;
remote task control from an iPhone;
and an always-on configuration.

It requires macOS 14 Sonoma or later. A Mac mini is recommended for continuous use but is not required. As of July 24, 2026, there was no Windows or Linux Personal Computer version.

Cloud Computer versus Personal Computer

	Computer	Personal Computer
Execution	Perplexity cloud	Native Mac plus Computer capabilities
Local files	Through uploads or connectors	Approved local folders
Native Mac apps	No direct desktop control	Supported
Always on	Cloud background	Can run on an always-on Mac
Platforms	Web, iOS, Android	macOS 14+
Risk focus	Cloud connectors and remote actions	Cloud risk plus local machine acces

Local access makes the approval model more important, not less.

Security and permission design

Computer operates in isolated cloud environments, while connectors provide authenticated access to real services. Enterprise controls can add administrative oversight.

Use least privilege:

connect a test workspace before production;
grant read-only scope when writing is unnecessary;
restrict folders, channels and repositories;
use a dedicated service identity where appropriate;
separate drafting from sending;
protect secrets and regulated data;
disconnect unused connectors;
audit recurring tasks and Skills.

Human checkpoints

Always require approval before:

sending an external communication;
publishing;
transferring money or buying;
deleting or overwriting data;
changing access or security;
accepting legal terms;
making an employment, medical, legal or financial decision;
or acting outside the stated system and scope.

How to pilot Computer

Week 1: read-only task

Choose a public or low-sensitivity workflow. Measure completion, verification time, credits and failure modes.

Week 2: private read access

Connect one bounded source. Confirm permission behavior and output handling.

Week 3: reversible write

Allow Computer to create a draft or write into a sandbox. Compare every change.

Week 4: scheduled or production trial

Add a narrow recurring task with alerts and a budget. Keep consequential action manual.

Define success before the pilot: time, quality, error, cost, employee effort and incident count.

Common mistakes

Giving a role instead of an outcome

“Be my chief of staff” is broad. Define one deliverable and authority boundary.

Connecting every app first

Connections should follow use cases, not precede them.

Hiding the final action in a long prompt

Put approval rules in a separate explicit section.

Scheduling before testing

An error repeated daily becomes an operational process.

Ignoring quiet failures

“No notification” may mean no change, an expired connector, insufficient credits or a task needing attention. Review run history.

Measuring only agent time

Count setup, supervision, verification, correction and incident handling.

Treating Brain as a database

Persistent AI context is not the system of record.

Frequently asked questions

Is Perplexity Computer the same as the Comet browser?

No. Comet is an AI browser. Computer is a broader cloud worker that can use web, files, connectors, subagents and schedules. Personal Computer includes a controllable Comet browser on Mac.

Is Computer included in Perplexity Pro?

Perplexity says active subscribers can access Computer. Pro does not have the same recurring credit allocation as Max, so usable capacity can depend on introductory or purchased credits.

How much do Computer credits cost?

Perplexity defines 100 credits as $1. Actual task cost varies.

Do normal Perplexity searches use credits?

No. Computer and selected premium features use credits; ordinary Ask searches do not.

What happens when credits reach zero?

Active work can pause. Scheduled runs can be skipped or retried and resume after capacity returns.

Can Computer send emails?

With an authorized connector, it can perform supported email work. Keep sending behind human confirmation until policy and testing permit otherwise.

Can Computer run automatically?

Yes. Scheduled Tasks support recurring work no more frequently than hourly.

What are Skills?

They are reusable procedures that guide Computer through a type of work and can be combined.

What is Brain?

It is a Max research-preview memory system intended to learn working context across projects, people and files.

Does Personal Computer require a Mac mini?

No. Any Mac running macOS 14 or later can use it. A Mac mini is recommended for always-on operation.

Is Personal Computer available for Windows?

Not as of July 24, 2026.

The bottom line

Perplexity Computer turns the answer engine into an action system. Its useful unit is not a message; it is a completed workflow across research, creation and tools.

That makes task design more important than clever prompting. Define the outcome, permitted context, approval gates, verification and budget. Begin with read-only work, measure real credit consumption and add automation only after the one-time task is dependable.

Computer can remove coordination overhead. It should not remove ownership.

byDavid Porter

Perplexity

Perplexity AI: Complete Guide to Search, Research, Comet and Computer

Perplexity Pricing Explained: Free vs Pro, Max, Enterprise and API

Write a comment