Cursor capabilities

Cursor's agentic surface includes the in-editor Agent mode (Composer), background/Cloud Agents that run on Anysphere VMs, Bugbot automated PR review, and Automations (scheduled or event-driven agents). Cursor 2.0 added parallel multi-agent execution (up to 8 agents on the same task via git worktrees, picking the best result).

Auto-Review is a Cursor agent run mode that lets the agent work longer with fewer manual approval prompts while still screening risky actions. Instead of the user approving every command (or flipping to a fully unattended mode), each agent tool call is routed through graduated trust levels: trusted calls run instantly, containable calls run in a sandbox, and everything else is screened by a classifier subagent that decides whether to allow it, try a different approach, or pause and ask the user for approval. Cursor describes the classifier as non-deterministic and best-effort convenience, not a security boundary.

Cursor Automations run cloud agents in the background on a schedule or in response to events, with no human at the keyboard. Triggers include recurring schedules (preset or a cron expression), GitHub/GitLab pull-request and CI events, Slack messages, Linear issues, PagerDuty incidents, Sentry errors, and custom inbound webhooks. When a trigger fires, an agent spins up an isolated cloud sandbox, follows your instructions using the MCP tools and models you configured, verifies its own output, and can draw on persistent Memories from prior runs. Typical uses: security review that audits diffs for vulnerabilities, bug triage (duplicate checks, issue creation, root-cause investigation), test-coverage gap reviews of merged code, incident response, and posting weekly codebase-change digests to Slack. Launched March 5, 2026.

Canvases let Cursor's agent respond with an interactive visual surface instead of a wall of chat text or a plain markdown file. The agent builds custom interfaces - sections, stats, tables, diagrams, charts, and data dashboards with logic and interactivity - so you can explore and digest results non-linearly. Canvases are durable artifacts that live in the Agents Window next to the terminal, browser, and source control, and are useful for reviewing PRs (logically grouping and prioritizing a change set), learning a new library, or managing other agents. Added in Cursor 3.1 as a visual surface beyond the diff view.

Cursor 3.9 (June 22, 2026) introduced the Customize page, a single hub replacing scattered settings screens for every extensibility surface: plugins (bundles of rules/skills/subagents/commands/MCP servers/hooks), rules, skills, subagents, hooks, slash commands, and MCP servers. It supports scoped management at the user, workspace, and team level, includes a marketplace leaderboard of the most popular plugins/skills/MCPs on the team, and ships plugin canvases (prebuilt, shareable setup templates such as a Hex canvas for data visualizations and an Atlassian canvas for live issues/projects/docs views).

Cursor edits one or many files in a single agent turn, presenting changes as inline diffs that the developer accepts or rejects per hunk or per file. The inline Cmd+K prompt edits a selection or empty buffer; Composer/Agent edits across many files at once.

Cursor is a standalone desktop IDE built as a fork of VS Code (Code OSS). It ships full editor functionality (file tree, terminals, debugger, source control) with AI features layered throughout, and is generally compatible with the VS Code extension ecosystem via Open VSX and direct VSIX install.

Conversational AI panel for asking questions about code, generating new code, and running multi-step agent tasks. Since Cursor 2.0 (Oct 2025) and Cursor 3.0, the historic Chat / Composer / Cmd+L panes were unified into an agent-centric 'Agents Window' that hosts chat, agents, PR review tabs, browsers, and terminals.

Cursor offers a model picker with multiple frontier models from Anthropic, OpenAI, Google, xAI, Moonshot, plus its in-house Composer family. Auto mode (and the Composer models) draw from a generous included 'Auto + Composer' usage pool; manually picking a third-party frontier model debits the separate API pool at provider token rates.

Six purchasable tiers (Hobby, Pro, Pro+, Ultra individual; Teams, Enterprise) — the marketing pricing page surfaces Hobby/Pro($20 'Individual')/Teams/Enterprise, with Pro+/Ultra detailed on the Models & Pricing docs. Paid plans now run a two-pool usage model: a generous 'Auto + Composer' pool (first-party Cursor models) plus a separate third-party 'API' pool with a monthly USD minimum (Pro $20, Pro+ $70, Ultra $400). Auto/Composer usage is heavily included; manually picking a frontier model debits the API pool at provider token rates. On-demand usage continues past the included pools, billed in arrears.

Cursor automatically builds a semantic index of the open workspace (code chunked into functions, classes, and logical blocks, then embedded as vectors) and exposes explicit context controls through @-mentions. The agent also performs agentic / semantic search at query time to pull in relevant files.

Cursor Tab is Anysphere's proprietary next-action prediction model offering inline single-token, multi-line, and cross-file edit suggestions. It anticipates the developer's next change (refactors, ripple edits) rather than only completing the current line.

Shipped June 30, 2026, this extends Cursor's team marketplace so admins configure an MCP server once from the Dashboard and distribute it across cloud agents, the Agents window, the IDE, and the CLI, instead of every team member configuring servers individually. Access can be restricted to specific organization groups, and existing SCIM directory group configurations are preserved when migrating prior Team MCP setups into the new marketplace model.

Cursor embeds the standard VS Code integrated terminal and adds an AI prompt for generating shell commands from natural language. The Agent / Composer can also execute terminal commands as part of multi-step tasks, governed by run modes (allowlisted commands run instantly, sandboxable ones run in a sandbox, everything else asks for approval).