Ask a developer what is Cursor and you will usually get a two year old answer: a fork of a popular code editor with a good autocomplete and a chat sidebar. That description is now wrong in a way that matters. Cursor’s own homepage calls the product "your coding agent for building ambitious software," and the official documentation opens by calling it "a coding agent for building ambitious software." The editor is still there, but it has been demoted to one client among several. So what is Cursor today? Answering that changes which tools you should compare it against.
The product changed underneath the name
The clearest evidence that Cursor is no longer defined by its editor is that Cursor now runs inside other people’s editors. On March 4, 2026 the company shipped Cursor agents into JetBrains IDEs through the Agent Client Protocol, covering IntelliJ IDEA, PyCharm, WebStorm and others. JetBrains’ own Aleksey Stukalov framed it as "Cursor becoming a special guest in the family of ACP-compliant agents in JetBrains IDEs," a setup where "developers stay in control of their environment, while Cursor brings the powerful AI assistance." A company whose moat was its editor does not ship that agent into a rival’s editor.
The same pattern shows up everywhere else. There is a terminal client, a web client at cursor.com/agents, an iOS app, iPad support added on July 29, 2026, a Slack bot, and comment-triggered agents in GitHub, Bitbucket and Linear. The JetBrains move is the tell; the surface area is the proof. Cursor sells an agent that happens to ship with an editor, not an editor that happens to include an agent.
What is Cursor today, surface by surface
Asking what is Cursor now means enumerating surfaces, not features. The desktop app remains the flagship. You open Agent with Cmd I or Ctrl I, and Shift+Tab toggles Plan Mode, which the docs describe as researching your codebase, asking clarifying questions and creating "a detailed implementation plan" before writing anything. Agent itself is documented as able to "complete complex coding tasks independently, run terminal commands, and edit code," with file search, web search, terminal execution, browser control and image generation among its tools. Notably, the docs state that "there is no limit on the number of tool calls Agent can make during a task."
Cloud Agents are the second surface and the more strategically interesting one. They run "in isolated VMs in the cloud with full development environments instead of on your local machine," complete with cloned repos, installed dependencies, secrets and network access. The Cloud Agents documentation is blunt about the intended usage pattern: "You can run as many agents as you want in parallel, and they do not require your local machine to be connected to the internet." They can be started from seven places, including a @cursor mention on a pull request.
Third is the CLI, which "lets you interact with AI agents directly from your terminal to write, review, and modify code," supports a print mode for CI pipelines, and can hand a conversation off to a Cloud Agent. Fourth is Bugbot, which "reviews pull requests and identifies bugs, security issues, and code quality problems" across GitHub, GitLab and Bitbucket, including self-hosted and Enterprise Server deployments.
Composer is a model now, not a panel
A lot of coverage still treats Composer as the multi-file editing pane. It is not. Composer is Cursor’s in-house model family. Composer 2.5 shipped on May 18, 2026, described in Cursor’s announcement as "a substantial improvement in intelligence and behavior over Composer 2," trained on "25x more synthetic tasks" than its predecessor using targeted reinforcement learning with textual feedback. The model card lists a 200k token context window and pricing of $0.50 per million input tokens, $0.20 per million cached input tokens and $2.50 per million output tokens, with a Fast variant at $3 and $15.
Cursor also co-owns a second model. The docs describe Grok 4.5 as "jointly trained by Cursor and SpaceXAI for long-running coding and knowledge work," priced at $2 per million input and $6 per million output tokens. That is a structural difference from tools that only broker somebody else’s models.
Model choice, Auto modes and the Router
Model selection in Cursor spans four vendors. Alongside Composer 2.5 and Grok 4.5, the models documentation lists Claude Sonnet 5, Opus 5 and Fable 5 from Anthropic, Gemini 3.1 Pro and 3.6 Flash from Google, and GPT-5.6 Luna, Sol and Terra from OpenAI.
You do not have to pick manually. Auto comes in three flavors. Auto Cost bills at a flat $1.25 per million input and $6 per million output tokens regardless of which model runs. Auto Balance and Auto Intelligence are "charged at Model API rates for the model used, based on actual usage." On Teams and Enterprise plans, Cursor Router, launched July 22, 2026, "picks the model for each Auto request based on your optimization mode." Admins toggle routing per team and per organization group, can disable up to two of the three modes, and can force Auto as the default picker. Where a model is blocked, the router routes "to an allowed model instead."
This is what separates Cursor from single-vendor agents. Weighed against Claude Code or OpenAI Codex, the routing layer is the real differentiator, not the editor chrome. Our head-to-head on those two covers that trade.
How pricing and usage limits actually work
The "500 fast requests" framing that dominated Cursor coverage in 2024 and 2025 is dead. According to the plans and pricing docs, paid tiers now run on two separate pools that reset monthly: a Cursor Models pool covering Grok 4.5 and Composer 2.5 with what the docs call "Generous included usage," and an Other Models pool for third-party models charged at API rates with a dollar-denominated monthly inclusion.
Pro is $20 per month and includes $20 of Other Models usage. Pro Plus is $60 and includes $70. Ultra is $200 and includes $400. A regional Start plan launched in India on July 28, 2026 at ₹649 per month, limited to Cursor models with fixed medium effort on Grok 4.5 and no fast mode. On the team side, Standard is $40 per user per month and Premium is $120 per user per month for five times the Agent limits. Hobby remains free with limited Agent requests and access to Composer.
When you exhaust a pool you can add on-demand usage and "continue at the same API rates with pay-as-you-go billing," or upgrade. The docs are explicit that request quality does not degrade near a limit. The consequence: on higher tiers you are buying wholesale API credit at a discount, so the sticker price is a poor guide to real spend once agents run in parallel.
Where Cursor genuinely differs from a terminal agent
Strip away the marketing and three differences survive.
The first is the extension supply chain. Cursor’s editor uses "the Open VSX extension registry for third-party extensions," but routes search and downloads through its own proxy at marketplace.cursorapi.com, running automated malware and supply-chain analysis first. The extensions documentation also warns that "the same publisher.extension name can point to different publishers or code on Open VSX than on the Microsoft Marketplace," and Anysphere publishes first-party replacements for popular extensions missing from Open VSX. A terminal agent has no equivalent problem or control.
The second is managed parallelism. Running ten terminal agents means ten sandboxes you built and maintain. Cursor’s Cloud Agents ship the environment, and the mobile and Slack entry points let you start and steer work without a laptop. It is the same shift toward cloud-hosted agent sessions seen elsewhere, from a different starting point.
The third is org-level policy. Enterprise admins can "set role permissions, whitelist or blocklist repos, models, and MCP servers, and configure global agent run settings." That last clause matters, because MCP servers are how agents reach production systems, raising the governance questions we covered in our piece on agent permissions in a CMS. Customization is layered: rules, skills, hooks, subagents where each one "operates in its own context window," and plugins bundling all of those together. Plugins launched February 17, 2026 with partners including AWS, Figma, Linear and Stripe.
Privacy, security and the enterprise posture
Privacy Mode is the load-bearing control. With it enabled, Cursor’s data use page states that "Customer Data will not be used for training by Cursor," backed by zero data retention agreements with providers. The caveat is stated plainly: providers may still "run risk classifiers to detect violations of terms and usage policies," and flagged data can be stored temporarily before deletion.
Codebase indexing deserves a closer read. Code is uploaded in chunks to compute embeddings, and "all plaintext code for computing embeddings ceases to exist after the life of the request." However, "the embeddings and metadata about your codebase (hashes, file names) may be stored in our database." Filenames and structure are not nothing in a regulated environment.
On the compliance side, the security page cites SOC 2 Type II attestation available on request through trust.cursor.com, commits to "at-least-annual penetration testing by reputable third parties," and states there are no operations in China and no Chinese-headquartered subprocessors. Enterprise adds SSO, SCIM provisioning, MDM deployment, audit logging and customer-managed encryption keys.
Frequently Asked Questions
Is Cursor still a fork of VS Code?
The desktop editor is still a fork, so keybindings feel familiar. But it now pulls extensions from Open VSX through Cursor’s own proxy rather than the Microsoft Marketplace, and the agent also runs in JetBrains IDEs, the terminal, Slack and the browser.
What is Cursor’s free tier good for?
The Hobby plan requires no credit card and gives you limited Agent requests plus access to Composer. It is enough to evaluate the workflow on a small project, not to run parallel agents or sustained long tasks.
Is Composer a feature or a model?
Both names have been used, but today Composer is Cursor’s in-house model family. Composer 2.5 launched May 18, 2026 with a 200k token context window and is billed from the Cursor Models usage pool rather than the third-party pool.
How do Cursor’s usage limits work now?
Paid plans include two monthly pools: Cursor Models and Other Models. Pro includes $20 of Other Models usage, Pro Plus $70 and Ultra $400. When a pool runs out you can add on-demand usage at the same API rates or upgrade your plan.
Which models can I use in Cursor?
Cursor lists Composer 2.5 and Grok 4.5 as its own models, plus Claude Sonnet 5, Opus 5 and Fable 5, Gemini 3.1 Pro and 3.6 Flash, and GPT-5.6 Luna, Sol and Terra. Auto and Cursor Router can select for you.
Does Cursor train on my code?
Not when Privacy Mode is on: the data use page says Customer Data will not be used for training by Cursor, and the company holds zero data retention agreements with model providers. Codebase embeddings and metadata such as hashes and file names may still be stored.
Can Cursor agents run without my machine?
Yes. Cloud Agents run in isolated cloud VMs with cloned repos, dependencies and network access, and the docs state they do not require your local machine to be online. You can start them from iOS, the web, Slack, GitHub or the API.
Are Cursor’s benchmark charts trustworthy?
Treat them as vendor claims until replicated. Cursor does disclose limitations, including that an earlier snapshot of its own codebase was accidentally included in Grok 4.5’s training, giving that model an advantage on the internal CursorBench evaluation.
