Cursor Essentials: Tab, Agent, and the Model Picker

Cursor is an AI code editor: a VS Code fork (built by Anysphere) with an agent in the core. Four surfaces do almost all the work — Tab (inline completion), Chat / Agent (Cmd+I), the mode picker (Shift+Tab), and the model picker. This guide gives the exact keybindings, modes, model names, config files, and prices to operate each, current as of 2026-06-15 (Cursor 3.7).

The 3.x Agents Window

On April 2, 2026, Cursor 3.0 introduced the Agents Window — a surface "centered around agents, while keeping the depth of a development environment." Open it with Cmd+Shift+P → Agents Window; toggle back to the classic IDE at any time. 3.0 added running many agents in parallel "locally, in worktrees, in the cloud, and on remote SSH," plus the /worktree (isolated changes) and /best-of-n (same task in parallel across models, each in its own worktree) commands.

Naming trap: in pre-3.0 tutorials "Composer" named the multi-file chat panel. In current Cursor the panel is just Agent (or Chat); Composer now names Cursor's own family of models (Composer 1 → 2.5).

Tab: completion that rewrites

Tab reads your recent edits, surrounding code, and linter errors, then proposes changes that can modify multiple lines, add missing imports, and coordinate edits across related code. Jump-in-file: after you accept, pressing Tab again predicts where your next edit belongs and jumps the cursor there. Cross-file suggestions: Tab can predict an edit in a different file and surface it in a "portal" window. Tab runs on a custom model Cursor trains itself (the docs page does not name the production model); per Cursor's engineering blog the next-generation model is Fusion, the system runs 400M+ predictions/day, and it is improved with online reinforcement learning on live accept/reject feedback.

Chat & Agent

Open with Cmd+I. The Agent "can complete complex coding tasks independently, run terminal commands, and edit code." It is built from three parts: Instructions (system prompt + your rules), Tools, and a user-selected Model.

Context with @-symbols

Type @ in the chat box to attach context instead of pasting it. You can also drop in images and voice, and switch models mid-chat without losing the thread.

Modes: Agent, Plan, Ask

The same chat surface runs in three modes; cycle them with Shift+Tab (or the mode-picker dropdown). The mode decides whether the agent edits immediately, plans first, or cannot edit at all.

• Agent — standard mode; goes straight to implementing changes. Use when you can describe the change precisely.
• Plan — "creates detailed implementation plans before writing any code." Asks clarifying questions, researches the codebase, and shows an editable plan before building. Use for complex, multi-file, or architectural work.
• Ask — read-only; answers questions and searches the codebase without editing (can use a read-only terminal, e.g. to inspect git history). Use to understand unfamiliar code first.

Model picker: Auto, Composer, frontier

The picker selects the model that drives the Agent and exposes three kinds of choice:

• Auto — "allows Cursor to select models that balance intelligence, cost efficiency, and reliability." It is a router, not a single model, and draws from the flat Auto + Composer usage pool. Token rates: $1.25/1M input, $6.00/1M output, $0.25/1M cache read.
• Composer (Cursor's own) — Composer 1, 1.5, 2, 2.5. A mixture-of-experts model RL-trained for agentic coding; Cursor positions it as a "Fast Frontier" model with generation speed ~4× faster than similar models. Cursor's blog lists GPT-5 and Sonnet 4.5 as "Best Frontier" models that outperform Composer on capability — the trade is speed vs. raw capability.
• Frontier models — five external providers (table below). The roster churns weekly; treat it as a 2026-06-15 snapshot and confirm on the live models page.

Rules: keeping the agent on-style

Cursor reads project rules so the agent follows your conventions automatically. Rules live in .cursor/rules/ as .mdc files (version-controlled; can nest in subfolders). Each file's frontmatter has three keys — alwaysApply (bool), description (string), globs (file-pattern string) — and those decide when the rule fires. As a simpler alternative, Cursor also reads a single plain-markdown AGENTS.md at the project root (no frontmatter).