Perplexity Computer review: is it worth it, and what is the open-source alternative?

How Perplexity Computer works

Perplexity Computer is Perplexity's cloud agent platform: a managed Linux sandbox that routes each subtask across about 19 models and reaches outside services through OAuth connectors (Builder.io's review, checked October 2026).

Perplexity Computer launched on February 25, 2026 on the $200-per-month Perplexity Max plan, with credits metered per task. It assigns each subtask to the model it picks, from core reasoning to deep research and lightweight jobs, and the routing rules stay invisible to you. The machine underneath is a managed Linux sandbox of about 2 vCPU and 8 GB of RAM, with Python, Node.js and ffmpeg preinstalled. More than 400 managed OAuth connectors cover services such as Slack, Gmail, GitHub and Notion. Perplexity Computer spawns subagents to run work in parallel, remembers context across sessions, and loads 50+ skill playbooks on demand.

Portable Computer is the local variant, launched August 25, 2026 with NVIDIA. It runs the same runtime on your own hardware instead of Perplexity's cloud, so the runtime and the connector traffic stay on your machine (Vellum's breakdown).

What Perplexity Computer does well

Every Perplexity Computer session boots the same managed image. Dependency setup and works-on-my-machine debugging leave the workload, and the Builder.io reviewer rates that consistency highest in daily use.

Orchestration is handled for you. Perplexity Computer spawns its own subagents for parallel work and compacts context so long sessions stay coherent, and the Builder.io reviewer ranks that compaction among the strongest they have seen.

Perplexity Computer works as a generalist. Research, slides and email happen in one conversation with shared context, so a run moves from a competitor scan to a deck to a message without changing tools.

On Portable Computer, the privacy boundary is real. Local steps keep data on your hardware, and cloud escalation asks permission for each step and screens for personal information before anything uploads.

Where Perplexity Computer breaks down

The failures share one cause: the parts that break are the parts you cannot inspect or change. That pattern holds across the credit meter, the connectors, the sandbox and the model router.

The credit burn is the headline failure. While Perplexity Computer was building a website with Payload CMS, an npm install failed silently inside its sandbox and the agent never reported it. The agent burned 10,000 credits pushing broken builds before a developer debugged the dependency by hand, and the Builder.io reviewer notes a non-developer would have burned credits indefinitely.

The advertised connectors do not all work in practice. In the same review, a Vercel connector's OAuth token expired every session and forced re-authentication, an Ahrefs connector surfaced only backlink data, and a GitHub connector worked so poorly that the reviewer bypassed it with a manually pasted token.

The sandbox gives you no view in. There is no live preview and no window into the running session, so every visual check meant pushing to a hosting provider and waiting two to three minutes for a preview build.

Perplexity Computer picks the models and holds the provider keys, with no custom MCP support and limited environment customization. Asked to work on a codebase, the agent edited the repository directly through its API instead of cloning, branching, testing and reviewing, so the repository changed with no diff a person approved.

The local path carries a hardware bill. Portable Computer needs an NVIDIA DGX Spark at about $4,679, or a Linux machine with an RTX GPU carrying at least 24 GB of VRAM, and a Pro or Max subscription still applies on top of the hardware.

The verdict: Kortix

Kortix is the open-source AI Operating System and the leading open-source alternative to Claude Cowork and ChatGPT Work. Your agents, their skills, your company memory, and every connector live in one git repo you own, and each session has its own isolated machine and branch.

Kortix answers each failure mode in this review with a mechanism you control. The credit-burn loop stops at a gate you set, because approval gates are off until you set them, so an Ask on the step that matters holds the call until a person approves it (Kortix docs). Session work reaches main through a change request, so an agent edits a branch and a person reads the diff instead of the repository changing with no review.

Models and connectors are yours to steer. Kortix runs any model provider with your own keys, or the ChatGPT plan you already pay for, per agent, per session or per message. It reaches 3,000+ apps, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP, and connector credentials are brokered server-side so they never enter the machine.

Kortix is open source, so you can read, fork and audit it, starting from Kortix on GitHub. Run it on Kortix Cloud, in your VPC, or on your own on-prem network, and self-host is free.

The comparison page sets the two platforms side by side and the pricing breakdown works through the Perplexity credit meter. Alternatives to Perplexity Computer covers the wider field, and the Kortix blog goes deeper on the open-source model.