Cursor CLI: Headless, Terminal-Native Coding (2026)

The CLI is Cursor's coding agent without the editor window. It runs in a terminal, so you can script it or run it headlessly in CI. The graphical editor is still the place for interactive editing and visual diff review; the CLI is the place for automation and long, unattended runs.

Type agent in any shell to open the Cursor agent. From there a slash-command menu is your control surface: it exposes the models you can switch to, your skills, the built-in commands and any MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers you have wired up. A few prefixes change how a turn runs rather than naming a command, and those are where most of the power hides.

& is a distinct mode. Prefix your message with it and the CLI spins up a cloud agent that opens its own branch and works async, so you can keep typing locally while it runs somewhere else. ! drops to a shell command from inside the agent without leaving the session. /cursor jumps from the terminal into the Cursor IDE and ports the conversation across, so you can hand a started task to the editor for visual review.

Shift+Tab cycles the run mode without leaving the prompt. It rotates through auto-run-all-commands, plan mode, ask mode, debug mode and plain agent mode, so you can drop into plan before a big change or flip to ask when you only want answers. The same terminal also exposes slash commandsSaved multi-step prompts you trigger from the / menu, such as run tests, fix CI or open a pull request. Press Enter for the full definition. to create rules, skills and sub-agents on the spot, plus the MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. login flow.

Yes, and the CLI makes it cheap to do. Open a terminal tab per model, start an agent in each, and point all of them at the same directory. One runs on Cursor's in-house ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. model, another on a flagship model like GPT-5.x, a third on Opus 4.x. Give every tab the same brief and you have fanned out parallel agents producing competing implementations of the same thing.

You switch a tab's model from the slash menu (or by typing agent then model), and you can set its reasoning level there too. It is a fast way to A/B the models on real work instead of arguing about which is best in the abstract: name the outputs, let all three run, then read the diffs side by side and keep the one you trust. The editor can do the same through separate chats, but the terminal makes the parallelism obvious.

agent -p "..." runs the agent headlessly: no interactive UI, just a prompt in and a result out. That is the form you drop into CI jobs, git hooks and shell scripts where nothing is sitting at a keyboard to approve steps. Pair it with tight permissions and a review gate, the same way you would any unattended automation.

The Cursor agent is not bound to one editor. Through ACPAgent Client Protocol. An open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio. Press Enter for the full definition. it runs inside JetBrains IDEs such as IntelliJ and PyCharm, carrying the same modes and models you use in the terminal and in Cursor itself. If your team lives in JetBrains, the agent meets them there rather than asking them to switch editors.

ACPAgent Client Protocol. An open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio. Press Enter for the full definition. is the Agent Client ProtocolAn open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio via the Cursor CLI bridge. Press Enter for the full definition., an open standard for coding agents: any IDE that implements it can host any agent that implements it. The Cursor side of that is the CLI. The same binary that runs agent in your terminal also runs the ACP bridge, so the JetBrains and Android Studio integrations are the CLI wearing a different front-end. MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. config, slash commandsSaved multi-step prompts you trigger from the / menu, such as run tests, fix CI or open a pull request. Press Enter for the full definition., rules, skills, sub-agents and model selection all behave the same way. Learn the CLI once and you know the ACP surface.

Android Studio is the clearest example. Install the JetBrains AI Assistant pluginA Cursor marketplace package that bundles MCP servers and skills (sometimes sub-agents and hooks); one click installs all of it into your Cursor instance. Press Enter for the full definition., install the Cursor CLICursor's command line: the full agent, all modes and models, in the terminal and pipeable into scripts and CI. Press Enter for the full definition. locally (it runs the bridge, so you need the binary), then in the IDE's AI chat install the Cursor agent from the ACPAgent Client Protocol. An open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio. Press Enter for the full definition. registry and authenticate either in the chat or via the CLI. Auth and the model you set in the CLI propagate up. You get the full Cursor agent experience inside Android Studio without leaving the IDE.

Treat the CLI reference docs as the contract when you wire Cursor into CI or a custom tool. The high-level guide explains when the CLI is useful; the reference pages pin down install commands, print mode, output formats, auth, config, permissions and the shell shortcuts that tend to break scripts when they drift.

The CLI shows per-agent context consumption as a running figure at the bottom of the session. As a session fills its window, quality starts to slip well before the window is technically full.

Good fits: failure triage in CI, and longer refactors you want to start from a script and walk away from. Because it is terminal-native, it slots into existing automation, but keep a human review gate. The CLI raises what arrives at review, and merging stays a deliberate, human step.

Scope what the agent may do before you let it run unattended: confine it to a branch, set permissions and shell access deliberately and require review on its output. Treat higher autonomy as something earned with evidence on your codebase, not the starting point. Check the official CLI docs for the current commands, flags and permission model before you wire it into a pipeline.