Perplexity builds a local-first AI agent that keeps secrets on your desk

Pumping confidential corporate files and private chat histories into cloud-hosted LLMs has always felt like a necessary evil. You pay through the nose for API tokens, and your sensitive data crosses into someone else's datacenter with every single query.

Perplexity just launched Computer, a local-first agent designed specifically for its Perplexity Portable Computer. Instead of pinging cloud APIs for routine thinking, the entire stack—model, harness, memory, and task execution—runs completely on the user's physical machine.

Sensitive files never leave the box without explicit permission, and since local compute doesn't charge per token, inference is effectively free.

Don't force small models to wear giant shoes

Open-source local models like Qwen 3.8 27B and NVIDIA Nemotron 3.5 Lightning 30B have gotten ridiculously capable, especially when paired with local hardware like an NVIDIA DGX Spark. But putting them inside general-purpose harnesses like Pi or Hermes usually ends in frustration.

Those open-source frameworks were built for massive cloud models. They assume your LLM can swallow giant system prompts, manage dozens of MCP tools simultaneously, and keep its train of thought over massive context windows.

Perplexity found that while Qwen 3.8 27B boasts a 260K token context window, its reasoning starts degrading once things push past 100K tokens.

So Perplexity trimmed the fat. The Computer harness uses a bare-bones system prompt and keeps only core tools in memory. Specialized abilities—like data visualization or coding—are loaded and unloaded as modular skills on demand, while stale context is actively summarized.

They also tossed out context-heavy Model Context Protocol (MCP) definitions for routine tasks. Connectors like Gmail, GitHub, Outlook, and Google Calendar were converted into lightweight command-line tools.

Security got an overhaul too. While Pi and Hermes run user-permission commands directly by default, Computer isolates all tool calls inside a strict OS-level sandbox. If the sandbox isn't working, the harness refuses to run at all.

Putting numbers on the board

Running Qwen 3.8 27B on an NVIDIA DGX Spark, Perplexity pitted Computer against Pi and Hermes. The differences were stark.

On Perplexity's internal benchmark of 53 local knowledge-work tasks, Computer hit 82.6%, beating Pi at 77.6% and Hermes at 74.0%. When swapped with Perplexity's fine-tuned PPLX 27B, that score climbed to 85.4%.

On the BrowseComp web research evaluation across 1,266 tasks, Computer reached 66.7% accuracy, compared to 50.2% for Pi and 43.9% for Hermes. Because Computer is paired with Perplexity's Search as Code pipeline, it resolved tasks in an average of 402.1 seconds and 852k tokens. That is 61% faster than Hermes and uses 70% fewer tokens than Pi.

On multimodal document parsing (ParseBench-100), Computer scored 65.1%, crushing Hermes (34.6%) and Pi (13.9%), with its biggest lead coming from chart comprehension at 76.5%. Complex layout understanding, however, stayed low across the board at just 16.2%.

Why it matters

We think this is a pivotal blueprint for the future of personal AI. Rather than treating local models like inferior drop-ins for cloud monoliths, Perplexity designed the software around the actual physical limitations of edge silicon.

For enterprise workers, researchers, and privacy-conscious pros, this unlocks an agent that can chew through private spreadsheets, local code repositories, and legal drafts without creating an audit nightmare or a massive monthly API bill.

It also fires a warning shot across the open-source agent tooling space. Endless tool sprawl and bloated protocol schemas kill small models. Stripping MCP servers down into compact CLI utilities is a remarkably clever move that the rest of the ecosystem will likely copy.

That said, running this setup smoothly still relies on hefty local gear like the DGX Spark. True on-device autonomy is finally here in software, but until everyday laptops ship with workstation-grade unified memory, the hardware price tag remains the real gatekeeper.