Perplexity has added a Windows client for its Portable Computer agent, enabling it to run locally on NVIDIA GeForce RTX and RTX PRO GPUs that have at least 24 GB of VRAM. For engineers, SREs, and security professionals this means multi‑step AI tasks can be performed on‑premises without automatically sending data to the cloud, preserving confidentiality while still offering an optional cloud fallback.
Supported hardware and model stack
The Windows release targets systems equipped with RTX GPUs that meet the 24 GB VRAM threshold. The agent ships with a pre‑configured local model – Qwen 3.8 27B – that has been post‑trained for Perplexity Computer workflows and tuned for NVIDIA RTX acceleration. No separate model download or custom build is required; the app handles model placement and GPU optimization internally.
Integration points and workflow extensions
Portable Computer includes built‑in connectors for common productivity and development tools: Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub. These connectors let the agent read files, query emails, or interact with pull‑request metadata without leaving the local environment. Sample use cases described by the vendor include:
- Engineering: automatically enumerate open GitHub pull requests, categorize them, and flag outdated documentation for review.
- Finance: ingest brokerage statements, 1099s, and tax returns stored locally, trace fee‑generating holdings, and cite exact file locations.
- Start‑ups: analyse funnel export data to locate signup drop‑offs and post insights directly to a Slack channel.
Operational considerations
Deploying the agent is intentionally lightweight: the installer provisions the local model and configures GPU drivers as needed. Practitioners should verify driver compatibility with their RTX hardware and monitor GPU memory usage, especially when processing large document sets. Because the agent can invoke cloud models for “more advanced research and reasoning,” administrators must decide how to handle the consent prompt – either by training users to approve on demand or by scripting an approved workflow that includes explicit user confirmation.
Security posture
All data processed by the local model remains on the device unless the user authorizes a cloud off‑load. This design reduces the attack surface associated with transmitting sensitive files to external services. However, the connector ecosystem introduces credential handling that must be secured – API tokens for Outlook, GitHub, or Slack should be stored using OS‑level secret stores or encrypted configuration files. Auditing the consent dialogs for cloud fallback is also advisable to ensure no inadvertent data leakage occurs.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams can now prototype or adopt AI‑driven automation without building a full model‑serving stack. The primary actions are to validate GPU capacity, lock down connector credentials, and establish a policy around cloud‑offload consent. Monitoring GPU utilization and logging consent events will provide the visibility needed to keep the workflow both efficient and compliant.
