Perplexity Puts Computer Behind MCP, Opening 400+ Services to AI Clients
The endpoint turns Computer into an execution layer for compatible assistants, but clients must preserve task state and surface logins, questions and sensitive-action approvals.
Listen to this story
The audio brief
Story brief
3 key pointsThe release gives MCP-compatible clients a way to delegate browser, coding, file-generation and business-system tasks to Perplexity Computer, not just query Perplexity’s APIs. Computer can reach more than 400 OAuth-connected services, but sensitive actions still require approval and tasks may take minutes. Developers must manage state through a reusable thread ID, resume after authentication or timeouts, and use...
- 01
The HTTP-based Computer MCP Server is distinct from Perplexity’s API MCP Server for Search and Completion.
- 02
Supported integrations include Gmail, Outlook, GitHub, Slack, Notion, Salesforce, Snowflake and Jira.
- 03
Sensitive actions return confirmation events; clients can approve or deny before execution continues.
Perplexity is putting its Computer agent behind the Model Context Protocol, giving compatible AI assistants a way to hand off multi-step work rather than only return an answer. The capability list is broad, spanning browsers, code, files and connected services, but the handoff can pause for authentication, user input or approval.
The Computer MCP Server is available through an HTTP endpoint for MCP clients that support HTTP transport. It is separate from Perplexity’s API MCP Server, which is focused on the company’s Search and Completion APIs. In this version, the client sends a task to Computer and receives structured results.
A browser, a sandbox and connected work systems
Computer can launch a cloud browser to navigate sites, fill forms and access login-gated pages. It can also run Python, Node.js and Unix tools in a sandboxed Linux virtual machine with a persistent workspace, then create PDFs, slide decks, spreadsheets, images and videos.
- Communication tasks can include reading and sending email through Gmail and Outlook, plus scheduling calendar events.
- Its supported project and knowledge connections include GitHub, Linear, Jira, Asana, Slack and Notion.
- Data work can draw on Snowflake, Databricks, Google Sheets and Salesforce.
The workflow has to survive interruptions
The exchange is stateful. The first Computer call returns a thread ID, which the client must reuse to continue the same workflow with its prior context. If Computer needs a service connection mid-task, it returns an authorization-required event; after the user completes the OAuth flow, the client resumes that thread.
Browsing, connector authentication, code execution and file generation can take several minutes. The service sends task-status progress notifications when a client supplies a progress token and supports them. A timeout can mean Computer missed the server’s wait window rather than stopped; the client can call the same thread again for status or continuation.
Account credits and direct uploads set the operating terms
The server uses OAuth 2.0 with a Perplexity account and does not require an API key. Usage is charged against that account’s credits; an insufficient balance produces an insufficient_credits response with a billing-management link.
Attachments follow a separate path: the client requests a presigned S3 upload envelope, uploads the file directly to S3, then passes its resulting URL to a Computer task. Perplexity says the MCP server never receives the file contents, and attachments can be up to 200 MB each.
Sources
- docs.perplexity.aiPerplexity Computer MCP Server - Perplexity