Cursor vs Claude Code

Which should you choose?

Start from where you want to spend the day, not from which model leads this month: both can run Anthropic’s models.

Choose Cursor if

• you want to read and change code in an editor, with the agent working beside you and its changes shown where you are already looking;
• you want models from several vendors on one bill: Claude, GPT, Gemini, Grok and Cursor’s own Composer;
• you want a free way in: Cursor’s Hobby plan includes a limited number of agent requests, while Claude Code is not on any free plan;
• you want to start cloud agents from Slack, Linear, GitHub or your phone.

Choose Claude Code if

• you live in a terminal, or move between editors, and want the same agent everywhere;
• you hand the agent long tasks across many files and come back to the result;
• you already pay for Claude Pro, Max or Team: Claude Code is included, starting on Opus 5.5, and its usage comes from your plan rather than from dollars at API prices;
• you would rather your prompts and code went to Anthropic alone, under Anthropic’s data terms.

The usual verdict in the comparisons we read is that Cursor is better inside the editor, Claude Code is better for autonomous work across many files, and many developers use both. That matches the shape of the two products, but no independent benchmark of the current versions backs either half of it.

Are Cursor and Claude Code the same?

No. Cursor is an editor with an agent inside; Claude Code is an agent that can sit inside an editor. The difference is where you start, and what is built around the agent.

Cursor is “based upon the VS Code codebase” and installs extensions from the Open VSX registry. Version 3.24 runs on macOS, Windows and Linux, and the same agent can be reached from the web, iOS, Slack, GitHub and Linear. Its command-line agent is called agent: it has Agent, Plan and Ask modes, a print mode (-p) for scripts, resumes with agent resume or --continue, and hands a task to a cloud agent with &.

Claude Code starts in the terminal: you run claude in a project folder and work with it there. The same agent comes as an extension for VS Code and Cursor, a JetBrains plugin, a desktop app for macOS and Windows, and on the web at claude.ai/code. It has no editor of its own, so you review its changes wherever you already review code.

# Cursor's CLI
curl https://cursor.com/install -fsS | bash

# Claude Code
curl -fsSL https://claude.ai/install.sh | bash

More install options, Homebrew included, are in Anthropic’s setup guide.

How each finds its way around your code

Both now find code by searching for it as they work. Cursor’s documentation describes Instant Grep, a search index built and queried on your machine, plus an Explore subagent that runs many searches in its own context window. It says Cursor “does not upload file paths or code to build a search index, and it does not store embeddings of your codebase for search.” Claude Code finds files by pattern, searches them with regular expressions using its built-in tools, and can hand work to subagents whose searching stays out of the main conversation.

If a comparison tells you Cursor’s advantage is a vector database of your code, it is describing something Cursor’s current docs do not.

Cursor vs Claude Code at a glance

As of 12 October 2026, from each company’s own documentation and pricing pages:

Price: dollars of usage against a plan allowance

Both start at $20 a month, but they meter differently, and neither publishes enough to compare the cost of a task. Cursor counts dollars of usage at each model’s API price; Claude counts an allowance it puts no numbers on.

Cursor

Cursor has a free Hobby plan with limited agent requests, and paid plans from Pro at $20 a month. Each plan includes two monthly pools: Cursor Models (Grok and Cursor’s own Composer, with what Cursor calls “significantly more included usage”) and Other Models (third-party models such as Claude and GPT, at each model’s API price). When a pool runs out, you pay on demand at the same rates or upgrade. Cursor does not publish how many dollars of usage each plan includes; its own guidance is that people who use the agent every day typically spend $60 to $100 a month. Current plans: Cursor’s pricing page.

Claude Code

Claude Code comes with every paid Claude plan, from Pro at $20 a month; it is not on the free plan. Use is metered in a rolling five-hour window with a weekly limit on top, shared with your chats on claude.ai, and Anthropic publishes no hour or prompt numbers. You can also pay per token with an API key. Current plans: Anthropic’s pricing page.

Models: many vendors against Claude only

Cursor offers models from most of the big labs; Claude Code runs Claude. If you want to switch vendors in the middle of a project, that alone decides it.

Cursor’s list, as of 12 October 2026, includes Anthropic’s Opus 5.5, Sonnet 5.5 and Fable; OpenAI’s GPT models; Google’s Gemini 3.1 Pro and 3.8 Flash; Grok; its own Composer 2.5; and open-weight models such as GLM and Kimi. The models Cursor now counts as its own, with the larger included allowance, are Grok (4.5 to 4.7) and Composer. One catch: Anthropic’s Fable needs data-retention approval, so with Privacy Mode on, Fable requests fail until an admin approves it.

Claude Code uses Claude models, from Anthropic directly or through Amazon Bedrock, Google Cloud or Microsoft Foundry. It starts on Opus 5.5 on every Claude plan (Pro and Team Standard moved up from Sonnet on 22 September 2026), and takes --model with aliases such as opus, sonnet and fable, plus effort levels from low to max.

The same model can behave differently in each, because the agent around it, with its own tools, prompts and context handling, is different. We know of no independent test of Opus 5.5 in Cursor against Opus 5.5 in Claude Code. If what you want is many vendors in a terminal agent, the open-source option is OpenCode; see OpenCode vs Claude Code.

Autonomy and approvals

Both now start with a classifier deciding what runs without asking you. Cursor’s recommended setup also sandboxes shell commands; Claude Code’s sandbox is opt-in.

Cursor’s run modes

Cursor has three run modes for local agents, under Settings › Agents › Approvals & Execution:

• Auto-review, the recommended default since Cursor 3.6 (29 May 2026). Allowlisted calls run at once, other shell commands run in a sandbox when they can, and anything else goes to a classifier model that runs on Cursor’s servers. It covers shell commands, MCP tools and web fetches.
• Allowlist: only what you have allowed runs without asking, and the sandbox is optional. An empty allowlist gives you the old Ask Every Time mode, which new users can no longer pick.
• Run Everything: every tool call runs automatically.

Cursor is unusually direct about its classifier: “Auto-review is not a security boundary. The classifier can make mistakes. It can allow a call you would have blocked, or block a call you would have allowed.” The sandbox blocks the network apart from an allowlist (by default yours plus common package registries), keeps writes inside the workspace and temporary folders, and protects paths such as .git/config and .git/hooks. Reads are looser: by default the agent can read files outside the workspace without asking, unless you switch Read Access to Workspace (added on 1 October 2026). Cloud Agents do not use run modes, and the CLI needs --force or --yolo to write files in a headless run.

Claude Code’s permission modes

Claude Code has started new interactive terminal and VS Code sessions in auto mode since late September 2026: a classifier reviews each action in the background and blocks the risky ones instead of asking you. Shift+Tab cycles to Manual (asks before edits and commands), Accept edits and Plan. dontAsk refuses anything that would prompt, and --dangerously-skip-permissions bypasses everything and is meant for disposable containers. Deny rules apply in every mode.

Its operating-system sandbox (Seatbelt on macOS) is off until you run /sandbox, and covers Bash commands only, not file edits, MCP tools or hooks. How to choose a mode safely is in Claude Code permissions, explained.

Rules files: .cursor/rules, AGENTS.md and CLAUDE.md

One AGENTS.md at the root of your repository now reaches both. Cursor’s editor and CLI read it, and Claude Code reads it when there is no CLAUDE.md.

Cursor’s own format is .cursor/rules/*.mdc: each rule can always apply, apply to files matching a glob, or be picked by the agent from its description. On top of those sit User Rules, Team Rules and AGENTS.md, at the root or nested in subfolders. Cursor’s CLI documentation says the CLI also reads CLAUDE.md at the project root; the editor’s rules page mentions AGENTS.md but not CLAUDE.md.

Claude Code reads CLAUDE.md from ~/.claude/, from the project (./CLAUDE.md or .claude/CLAUDE.md) and from a personal CLAUDE.local.md, plus rule files in .claude/rules/. Since version 2.1.277 (18 September 2026) it reads AGENTS.md natively when a project has no CLAUDE.md.

MCP and extensions

Both connect to MCP servers, but each keeps its own config file, so you set servers up twice. Cursor reads .cursor/mcp.json in the project and ~/.cursor/mcp.json for you, over stdio, SSE or Streamable HTTP with OAuth, and supports MCP tools, prompts, resources, roots and elicitation; its CLI uses the same files. Claude Code reads a project’s .mcp.json and your ~/.claude.json.

Around that, Claude Code has hooks, skills, plugins with marketplaces, and subagents (more in Claude Code hooks, explained). Cursor has the VS Code extension ecosystem through Open VSX, and Anthropic’s Claude Code extension is one of the things you can install there.

Does Cursor collect your data? Privacy compared

Cursor sends the code its agent reads to the model you chose, as every cloud AI tool does. Whether it can be used for training depends on Privacy Mode, which is on by default only for Enterprise. On any other plan, check the setting yourself.

With Privacy Mode on, Cursor says your code is never used for training and its model providers keep none of it (zero data retention). Its search index stays on your machine, but when the agent opens a file, “that file content can still be included in the model request.” Cloud Agents keep an encrypted copy of your repository while they run.

Claude Code runs on your machine and sends your prompts, including the code it reads, and the model’s replies to Anthropic over TLS. On Pro and Max, Anthropic trains on that data only if your privacy setting allows it, and keeps it for five years if it does, 30 days if not. On Team, Enterprise and the API, it is not used for training and is kept for 30 days as standard; Enterprise organisations can qualify for zero data retention. Claude Code’s usage metrics never include code, prompts or file paths, and DISABLE_TELEMETRY=1 turns them off. Session transcripts stay on your machine, in plain text under ~/.claude/projects/, for 30 days by default.

Cloud agents and working away from your desk

Both can run agents on cloud machines as well as on yours. Cursor’s start from more places and bill at API prices; Claude Code’s run from the web and the mobile app.

Cursor’s Cloud Agents, formerly Background Agents, run in isolated cloud VMs and can be started from the desktop app, the web, iOS, Slack, GitHub or Bitbucket, Linear and an API. They are billed “at API pricing for the selected model”. September added self-hosted machines (2 September), Projects with coordinator agents (10 September), and Rollouts with Security Review (23 September); since 6 October the iOS app can also control agents running on your own computer.

Claude Code’s cloud sessions run in Anthropic-managed VMs, or on infrastructure your organisation runs, from claude.ai/code and the mobile app. A session can move between your terminal and the cloud (--cloud, --teleport), Remote Control drives a local session from a browser, and Routines (/schedule) run tasks on a timetable. On GitHub, Claude answers @claude mentions through its action.

SpaceX now owns Cursor

SpaceX agreed to buy Cursor on 16 June 2026, in an all-stock deal CNBC valued at $60 billion, and Cursor announced on 14 August that it had closed.

The visible change so far is in pricing: the models Cursor counts as its own, with the larger included allowance, are now Grok and Composer. Grok comes from SpaceXAI, SpaceX’s AI company. Cursor’s announcement says the acquisition grew out of an April partnership with SpaceXAI to train models, and calls Grok 4.6, released two days before the close, “an early look at what we can now build together”.

As of 12 October, Cursor’s model list still includes Claude, GPT and Gemini, and we found no public commitment from Cursor either way about keeping third-party models. If your work depends on one vendor’s models inside Cursor, that is worth knowing; it is not evidence that anything will change.

What other comparisons miss

Most of the comparisons ranking for this search describe a Cursor and a Claude Code that no longer exist. Of the top results we could read on 12 October 2026:

• none mentioned that SpaceX owns Cursor, or that Cursor’s included usage now favours Grok and Composer;
• none described Cursor’s current pricing, with two usage pools, Pro Plus at $60 and API prices for third-party models; one updated in September 2026 still quotes “~500 fast requests”, Max Mode (now only on legacy plans) and Claude 3.5 Sonnet;
• none mentioned Cursor’s agent CLI, or that it reads CLAUDE.md;
• none mentioned that Claude Code’s own extension installs in Cursor, which makes “Cursor or Claude Code” a false choice for many people;
• one says Cursor has no native terminal execution layer, but its agent runs shell commands, in a sandbox under its recommended setup;
• one of the top results compares the Claude chat app, not Claude Code.

Where each falls short

Cursor

• It does not publish how much usage each plan includes, and agent work on third-party models is billed at API prices.
• Privacy Mode is on by default only for Enterprise.
• By its own account, Auto-review “is not a security boundary”, and by default the agent may read files outside your project without asking.
• Its larger included allowance now goes to its owner’s models, with no public commitment about third-party ones.

Claude Code

• Claude models only.
• No free plan, and no published limit numbers: you find the edge of your plan by reaching it.
• The OS sandbox is off by default and covers Bash only.
• No editor of its own: you review its changes in another tool.
• In our testing, resume is tied to the folder: claude -c and claude --resume only find sessions started in the same working directory. And its hooks report a permission prompt, but nothing when a turn is interrupted or fails with an API error.

Using both

Using both is common and simple, because Claude Code runs inside Cursor. Anthropic’s documentation has an “Install for Cursor” link for the Claude Code extension, which is also on Open VSX, and the CLI works in any terminal, Cursor’s included.

One way to split the work: Cursor’s editor and agent for changes you want to watch line by line, Claude Code for longer tasks you hand off, each billed to its own plan. Keep one AGENTS.md at the root so both follow the same rules, and add your MCP servers to both .cursor/mcp.json and .mcp.json.

Two cautions. A Claude model in Cursor’s agent and Claude Code in Cursor’s terminal are billed separately: the first by Cursor, the second by your Claude plan. And two agents editing the same working tree will trip over each other, so give each its own branch or git worktree, as in running Claude Code sessions in parallel.

Once Claude Code is one of several agents, knowing which one is waiting on you becomes the hard part; that is what we built AI Bridge for (see running several AI coding agents at once).

And Codex?

Codex sits on the same side of the line as Claude Code: a terminal agent, with an IDE extension that also installs in Cursor. OpenAI’s Codex CLI is open source (Apache-2.0), comes with ChatGPT Plus ($20 a month) and above, runs in an operating-system sandbox with the network off from its first run, and defaults to GPT-6.1 Sol.

So “Codex vs Cursor” is the same editor-or-agent question as this page, with OpenAI’s models and plans in place of Anthropic’s. How the two terminal agents compare with each other is in Codex vs Claude Code.

Questions