Perplexity AI launched a local version of its autonomous agent system called Portable Computer on Tuesday, giving users the ability to run automated AI tasks directly on their own desktop hardware instead of third-party cloud servers. Developed with Nvidia, the software lets users keep their private documents, spreadsheets, and source code on their personal machines while using an agent that reads files, executes shell commands, and connects with work apps. The system runs tasks locally without charging monthly cloud usage credits, and it asks user permission before it routes any difficult query to a remote cloud model.
Key Takeaways
- Portable Computer runs local AI models directly on user hardware, keeping private files and prompts on the device.
- Tasks processed on local hardware consume zero billing credits or API tokens.
- The app integrates an isolated operating system sandbox, agent tools, and connectors for Google Drive, Gmail, Slack, and GitHub.
- Users need high-end hardware, starting with Nvidia Linux systems with 24GB VRAM, with Windows support arriving in September.
How the Local Agent Works
Most popular AI chatbots send user text and uploaded files to cloud data centers for processing. Perplexity designed Portable Computer as a bundled application that contains the local model weights, the inference engine, and connector tools inside a single setup.
The agent operates in an isolated desktop sandbox. When an Indian software developer or financial analyst assigns a complex assignment, like reviewing tax forms or analyzing sales spreadsheets, Portable Computer opens the documents locally, extracts details, and writes summaries without uploading raw data to the web.
If a multi-step project requires external information or complex reasoning, the system employs hybrid routing. It completes standard data parsing locally, pauses to ask user permission, and then routes only that specific prompt to larger cloud models. Users only spend subscription tokens on the specific steps sent to the web.
Model Choices and Strict Hardware Needs
At launch, users can run the open-source Qwen 3.8 27B model or Perplexity’s fine-tuned variant called PPLX 27B. Perplexity also plans to add Nvidia’s Nemotron 3.5 Lightning model soon.
Running these models locally requires substantial desktop power. The software works on Linux machines equipped with Nvidia graphics cards, such as the Nvidia DGX Spark desktop supercomputer or systems running Ubuntu with at least 24GB of video memory, including GPUs like the RTX 3090, RTX 4090, or newer workstation cards. Perplexity plans to roll out Windows support next month. The tool is open to Perplexity Pro, Max, Enterprise Pro, and Enterprise Max subscribers.
Frequently Asked Questions
Q1. What is Perplexity Portable Computer?
A1. It is a desktop application that runs an autonomous AI agent locally on your own computer hardware instead of relying only on cloud servers.
Q2. Do I need an internet connection to use Portable Computer?
A2. You can run basic local tasks and document reviews offline, but features like web search, app syncing, and cloud model handoffs require internet access.
Q3. What hardware do I need to run it?
A3. You need an Nvidia graphics card with at least 24GB of VRAM running on Linux or DGX OS, with Windows PC compatibility scheduled for September.
Q4. Does running local tasks cost subscription credits?
A4. No. All processing done on your local GPU uses zero token credits, and you only spend credits when you approve sending a step to cloud models.
