How does "Perplexity Computer" work?

Perplexity’s “Computer” (desktop app with Computer mode and MCP support) runs Perplexity’s own models and tools locally on your machine, and can also route work to external models (OpenAI, Google, xAI, etc.) behind the scenes, but it does not “plug into” Opus/Gemini/Grok/Veo in the way you manually wire separate APIs together. Instead, Perplexity abstracts those models behind a single interface and lets you add your own tools and data via MCP and Connectors.

What “Computer” actually is

On macOS, Perplexity exposes Local MCP support inside the desktop app. This lets the AI talk to your local environment (files, apps, terminals, calendars, etc.) using the Model Context Protocol.

• You enable this by installing the PerplexityXPC helper and then adding “Connectors” pointing at MCP servers running on your machine.
• Once configured, Computer mode can call those MCP tools (e.g., “check my Mac calendar”, control iTerm, interact with local DBs) in the middle of a normal chat.

This is orthogonal to whether the underlying LLM is Perplexity’s own, GPT‑4/4.1, or any other third‑party model; the MCP layer sits on top as a tool interface.

How external LLMs are integrated

Perplexity runs a model router over multiple first‑party and third‑party models and chooses (or lets you choose) which model to answer a given query. The public “Third‑Party Models & Terms” page confirms that Perplexity integrates a set of external models from different providers, and which ones apply depends on your plan and settings.

• For text: the main chat experience uses a combination of Perplexity’s own models and third‑party LLMs, selected dynamically by plan, query type, and settings.
• For images: Perplexity exposes multiple image generators (GPT Image 1 from OpenAI, Google “Nano Banana”, Bytedance Seedream) behind a single “generate image” instruction in the UI; you can set a preferred one in Settings or leave it on “Default”.

Conceptually, Opus/Gemini/Grok/etc. would fit into this model pool: Perplexity’s router can call out to them as providers when licensed/available, but you as a user don’t wire them in directly with API keys inside Computer; they are abstracted behind the Perplexity platform.

Where MCP and third‑party models meet

MCP is about tools and data access, not which LLM is used.

• A given chat turn may use: A Perplexity‑ or third‑party LLM (e.g., an OpenAI or Google model) to reason and plan, or MCP tools (local or remote) to fetch or modify data (files, apps, SaaS) invoked by that LLM via the MCP protocol.
• From your perspective, you just issue a natural language request; Computer mode decides when to call MCP tools versus pure model reasoning, and Perplexity decides which underlying model to run that reasoning on.

So “integration with Opus, Gemini, Grok, Veo, etc.” is not done at the MCP/Computer configuration level; it happens at the platform level where Perplexity routes your request to the best available model and then optionally augments it with MCP and Connectors. You configure tools (MCP, Connectors) and preferences (e.g., image model); Perplexity manages the actual external LLM integrations behind the scenes.

If you tell me which specific model you care about (e.g., Gemini 2.0 vs Claude Opus vs “Grok 3” APIs), I can outline how you’d normally use that separately via its own API, and how MCP/Computer would sit alongside it in your workflow rather than directly “inside” it.

See their introduction for yourself -