Ask by product,
task, or condition.

ChatGPT can connect to custom apps built with MCP, with workspace controls and plan-dependent app capabilities.

ChatGPT exposes web search as a conversation tool, including inside Projects when the workspace allows it.

ChatGPT agent can use its own computer to complete multi-step web tasks under user supervision.

ChatGPT uses a Python code interpreter for structured data and other file-processing tasks on supported plans.

ChatGPT accepts documents, spreadsheets, presentations, images, and other supported files; limits depend on plan and surface.

ChatGPT analyzes PDF text, and ChatGPT Enterprise also documents visual retrieval for charts, diagrams, and embedded images.

ChatGPT analyzes CSV, XLS, and XLSX data through its code interpreter on supported plans.

ChatGPT accepts static image inputs on documented supported plans and web or mobile surfaces.

ChatGPT Images creates new images and edits uploaded or previously generated images; basic generation is documented across all tiers.

ChatGPT Projects retain chats, files, and instructions as project context, while broader memory availability depends on settings and workspace policy.

ChatGPT apps can search external data, run deep research, synchronize content, and in some cases take actions.

ChatGPT is documented across web and desktop workflows, with dedicated desktop applications for macOS and Windows.

Claude supports MCP across Claude.ai, Claude Desktop, Claude Code, and the Messages API, with setup varying by surface.

Claude Research uses live web search and returns cited research on paid Claude plans.

Claude provides a sandboxed computing environment for code execution, file creation, and data analysis.

Claude accepts documents and images in chats and projects, subject to file-size, count, and context limits.

Claude analyzes PDF text and, for PDFs under the documented page threshold, visual elements such as charts and graphics.

Claude accepts XLSX when code execution is enabled and can analyze CSV, TSV, and spreadsheet data in its computing environment.

Claude accepts JPEG, PNG, GIF, and WebP image uploads and can interpret supported visual content.

Claude can search and reuse context from previous conversations for paid-plan users where chat search and memory are available.

Claude supports remote connectors across its major surfaces and local desktop extensions on Claude Desktop and Claude Code.

Claude Desktop is documented for macOS, Windows, and Linux.

Perplexity supports custom remote connectors built with MCP in its enterprise connector system.

Perplexity combines live web search with uploaded or connected internal knowledge in supported search workflows.

Perplexity Computer can browse the web and act across connected services as a separate agent surface.

Perplexity Computer can create functional applications and execute tool-backed workflows in an isolated cloud environment.

Perplexity accepts common document and data files for internal knowledge search and project context on paid plans.

Perplexity supports PDF files in Internal Knowledge Search and file-backed research workflows.

Perplexity supports XLSX and CSV files in Internal Knowledge Search.

Perplexity accepts image attachments in individual queries, but its Internal Knowledge Search does not index PNG or JPEG files.

Perplexity Computer documents rich-media generation for visual content as part of its agent workflows.

Perplexity stores project-scoped memories, while Computer also documents persistent context across sessions.

Perplexity Computer can run asynchronous work and scheduled monitoring while the user is away.

Perplexity offers file, app, and custom remote connectors, with connector availability varying by plan.

Perplexity Computer is available on desktop web and mobile; this evidence does not establish a native desktop application.

Codex supports MCP for connecting external tools and systems.

Codex includes a web-search tool for current external information.

The Codex app includes an in-app browser that Codex can control for localhost web development, with broader control described as evolving.

Codex runs software-engineering tasks in command-capable sandbox environments and is also available as a terminal CLI.

Codex executes code and tests inside isolated task environments while working on a repository.

Codex can inspect screenshots, technical diagrams, charts, and other visual surfaces during coding work.

OpenAI documents image generation and editing as available inside Codex.

Codex cloud tasks run independently in separate environments and can be delegated in parallel.

Codex connects to external systems through configured MCP servers and other available tools.

OpenAI offers a Codex desktop app with an integrated workspace and browser.

Codex operates across app, CLI, and IDE extension surfaces, with shared agent workflows.

Claude Code can connect to MCP servers and can also expose Claude Code as an MCP server.

Claude Code runs as an interactive terminal agent and can execute approved shell workflows.

Claude Code can inspect repositories, edit files, and run commands from its terminal agent loop.

Claude Code can use remote connectors and local desktop extensions through Anthropic’s connector architecture.

Claude Desktop can deep-link directly into a Code session on macOS, Windows, and Linux.

Claude Code supports IDE-integrated workflows in addition to its command-line interface.

Cursor connects to local and remote MCP servers and provides a curated MCP server directory.

Cursor Agent includes web search and URL fetching among its research tools.

Cursor can control a browser through configured integrations such as the curated Playwright MCP server.

Cursor Agent can run commands in the editor terminal, and Cursor also provides a terminal-native CLI agent.

Cursor Agent can edit repository files and execute commands and tests through the terminal tool.

Cursor supports remote background agents launched from web or mobile and handed back to the desktop editor.

Cursor’s MCP directory provides connectors for services including GitHub, Linear, Notion, Figma, and databases.

Cursor is an installed desktop code editor for macOS, Windows, and Linux.

Cursor is itself an AI-first IDE and exposes its agent tools inside the editor.

GitHub Copilot supports MCP tools in CLI, IDE, and configured cloud-agent contexts, subject to enterprise policy.

GitHub Copilot cloud agents can use an out-of-the-box Playwright MCP server against localhost.

GitHub Copilot CLI runs as an agent in the terminal and can perform approved command-line work.

GitHub Copilot agents can make code changes and use command execution in terminal, IDE, and cloud workflows.

GitHub Copilot can delegate cloud coding work that produces changes and pull requests for later review.

GitHub Copilot can connect to external services through configured MCP servers and organization-approved registries.

GitHub Copilot integrates with VS Code, Visual Studio, JetBrains IDEs, Xcode, and other supported development environments.

Gemini Code Assist agent mode can use configured MCP servers; Google labels agent mode as preview.

Gemini Code Assist agent mode can use Google Search responses in VS Code context.

Gemini Code Assist agent mode exposes Gemini CLI built-in tools, including terminal access, in VS Code.

Gemini Code Assist agent mode can write files and run tools as part of multi-step coding work.

Gemini Code Assist can connect to external services through MCP servers and RESTful API tools in agent mode.

Gemini Code Assist agent mode is available inside VS Code and IntelliJ-family IDEs.

Devin Desktop’s Cascade agent natively connects to user-configured MCP servers.

Cascade includes a dedicated web-search tool among its built-in agent capabilities.

Cascade supports browser-control actions with configurable auto-execution permissions.

Cascade can run terminal commands and install required packages as part of agent work.

Cascade edits code and executes terminal workflows inside the development environment.

Devin Desktop exposes memories and rules for reusable project context; exact persistence depends on configuration.

Cascade connects to third-party systems through MCP servers, with additional enterprise controls.

Devin Desktop is an installed development environment for major desktop operating systems.

Cascade operates directly inside Devin Desktop’s code-editor environment.

Replit Agent can test a built application automatically in a real browser when App Testing is enabled.

Replit Agent can run build and debugging workflows inside Replit’s hosted development environment.

Replit Agent writes, runs, debugs, and improves applications in Replit’s hosted runtime.

Replit checkpoints preserve Agent conversation context, project state, environment configuration, and Agent memory.

Replit Agent supports extended autonomous runs, while availability and duration depend on selected mode and account limits.

Replit Agent can configure application integrations, but this review did not verify general MCP-client support.

Replit Agent runs primarily in Replit’s web and mobile cloud surfaces rather than a verified native desktop application.