Claude Code vs Cursor: how the two coding agents differ
Claude Code and Cursor are AI coding agents that start from opposite ends of the developer workflow: Claude Code is Anthropic's agent that runs in a terminal and reaches into editors, a desktop app, and the web, while Cursor is an AI code editor that added a CLI and cloud agents around the editor. Since Cursor can run the same Claude models, the Claude Code vs Cursor choice turns less on code quality and more on form factor, model access, pricing structure, and governance.
This comparison goes deeper on the pair than the broader AI coding agents comparison. Every product detail was checked against Anthropic's and Cursor's documentation, changelogs, and pricing pages on November 16, 2026. Both products change often, so treat plan names and limits as a dated snapshot.
Claude Code vs Cursor at a glance#
Claude Code is a model vendor's agent that runs on many surfaces, and Cursor is an editor vendor's agent that runs many vendors' models. The table summarizes the main differences as documented on November 16, 2026. The sections after it explain each row.
| Claude Code | Cursor | |
|---|---|---|
| Maker | Anthropic | Anysphere, acquired by SpaceX in August 2026 |
| Primary form | Terminal CLI | Desktop editor based on the VS Code codebase |
| Other surfaces | VS Code and JetBrains extensions, desktop app, web, mobile, Slack | Agents Window, CLI, web, iOS app, JetBrains via ACP, Slack, GitHub, Linear |
| Models | Anthropic's Claude models only | Cursor's Grok and Composer models, plus Anthropic, OpenAI, Google, and others |
| Model requests go to | Anthropic, Amazon Bedrock, Google Cloud, Microsoft Foundry, or your gateway | Cursor's backend, then the model provider |
| Instruction files | CLAUDE.md, or AGENTS.md when no CLAUDE.md exists |
.cursor/rules, AGENTS.md, and CLAUDE.md |
| Pull request review | Code Review (Team and Enterprise) | Bugbot (Premium and up) |
| Lowest paid plan | Claude Pro, $20 a month | Cursor Plus, $8 a month, Cursor models only |
What is Claude Code?#
Claude Code is Anthropic's agentic coding tool: it reads a codebase, edits files, runs commands, and connects to development tools from a terminal, IDE, desktop app, or browser.
Anthropic's Claude Code documentation describes one engine behind every surface, so a repository's CLAUDE.md files, settings, and MCP servers carry over between the CLI, the IDE extensions, and the desktop app. The agent gathers context, takes action, and verifies results in a loop until the task is done or it needs input. Claude Code is proprietary; version 2.1.325, released November 13, 2026, was the latest in its changelog on the day of this comparison.
What is Cursor?#
Cursor is an AI code editor and coding agent from Anysphere, based on the VS Code codebase, that runs agents in the editor, a terminal CLI, and cloud virtual machines.
Cursor 3, released April 2, 2026, made the Agents Window generally available: an agent-first workspace for supervising many local and cloud agents alongside the classic editor. Cursor's blog announced on August 14, 2026 that SpaceX had completed its acquisition of the company. Cursor now sells its plans as part of The X Subscription, which bundles Cursor with Grok and X, and runs Origin, a git forge in early beta.
Form factor: terminal agent vs AI-native editor#
The largest practical difference is where each product expects the developer to sit. Claude Code expects a task described in a terminal or chat panel and returns a diff to review. Cursor expects the developer in an editor, with agent panels beside the code, and adds a CLI and cloud agents around that editor.
Claude Code's surfaces. The CLI installs with one command (curl -fsSL https://claude.ai/install.sh | bash) and is the most complete surface; Anthropic's platform guide says scripting and the Agent SDK are CLI-only. The VS Code extension adds inline diffs and plan review, and Anthropic offers an install link for it in Cursor too. A JetBrains plugin, a desktop app for macOS and Windows (Linux in beta), cloud sessions at claude.ai/code that run in Anthropic-managed VMs after the laptop closes, the Claude mobile apps, and @Claude mentions in Slack complete the set.
Cursor's surfaces. The editor keeps VS Code's default shortcuts and extension model, with extensions from the Open VSX registry. Agent runs in a side panel and Tab offers inline completions. The Agents Window manages agents across workspaces and moves them between local and cloud. The CLI runs as agent, and prefixing a message with & hands it to a cloud agent. Cloud agents run in isolated VMs and start from the desktop app, cursor.com/agents, the iOS app, Slack, GitHub, Bitbucket, Linear, or an API. JetBrains users reach Cursor through the Agent Client Protocol.
Models: Claude only vs a multi-vendor catalog#
Claude Code runs only Anthropic's Claude models, served by Anthropic, a cloud provider, or a gateway. Cursor offers its own Grok and Composer models plus third-party models from Anthropic, OpenAI, Google, and others. If a team wants Claude, both tools provide it; if it wants to switch vendors per task, only Cursor does.
Claude Code. Anthropic's model configuration guide defines aliases that track the recommended version per provider. On the Anthropic API, opus resolved to Opus 5.5 (the default on every plan that includes Claude Code and on the API), sonnet to Sonnet 5.5, haiku to Haiku 5.5, and fable to Fable 5.2; opusplan plans with Opus and executes with Sonnet. Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5, and Fable models run with a 1 million token context window by default. Claude models are also available through Amazon Bedrock, Google Cloud's Agent Platform (formerly Vertex AI), and Microsoft Foundry, where aliases can resolve to older versions; on Bedrock, sonnet resolved to Sonnet 4.5. Anthropic states that it doesn't support routing Claude Code to non-Claude models through any gateway.
Cursor. Cursor's models and pricing page lists Grok 4.7, 4.6, and 4.5 and Composer 2.5 as Cursor models, describing the Grok models as jointly trained by Cursor and SpaceXAI. The third-party catalog includes Claude models through Opus 5.5, Sonnet 5.5, and Fable 5.2; OpenAI's GPT-5 and Codex models; Google's Gemini models; and models from DeepSeek, Z.ai, Moonshot, and Meta. Auto mode picks a model for cost, balance, or intelligence. Cursor listed Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, the same as Anthropic's API list price. Fable models require admin approval of Anthropic's data retention terms for Enterprise teams and Privacy Mode users, and Tab completions use Cursor's built-in models even when you add your own API keys.
A benchmark score does not settle the comparison. A SWE-bench result describes a model and an agent configuration together, and both tools can run the same Claude model, so what differs is the agent's tooling and context handling. The LLM benchmarks guide explains why vendor-reported coding scores need care.
Agent capabilities compared#
Feature for feature, the two products have converged. Both ship subagents, hooks, skills, plugins, MCP, plan modes, checkpoints, classifier-based approvals, pull request review, a headless mode, and SDKs. The differences are in defaults, file locations, and where each feature runs, as the table and the sections below show.
| Capability | Claude Code | Cursor |
|---|---|---|
| Context gathering | Agentic search with file, grep, and glob tools; MCP tool definitions loaded on demand | Instant Grep local index; @-mentions for files, terminals, and diffs |
| Planning | Plan mode; opusplan alias |
Plan mode; Projects with a coordinator agent |
| Subagents | Explore, Plan, general-purpose; custom in .claude/agents/; agent teams |
Explore, Browser; custom in .cursor/agents/ or .claude/agents/; cloud subagents |
| Hooks | Run a command, HTTP request, MCP tool, prompt, or subagent on lifecycle events | hooks.json for agent, Tab, and app events; loads Claude Code hooks |
| Skills | Agent Skills standard in .claude/skills/ |
Agent Skills in .cursor/skills/ and .agents/skills/; loads .claude/skills/ |
| Plugins | Plugins and marketplaces | Cursor Plugins, Agent Plugins standard, team marketplaces |
| MCP | Local and remote servers; managed allowlists | stdio, SSE, Streamable HTTP, OAuth; Enterprise allowlist |
| Checkpoints | Before each prompt; /rewind restores code, conversation, or both |
Before significant changes; restores files |
| Approvals | Permission modes, including classifier-based auto mode; optional Bash sandbox | Auto-review classifier, Allowlist, or Run Everything; shell sandbox |
| Inline completion | None documented | Tab |
| Headless and CI | claude -p; GitHub Actions; GitLab CI/CD |
agent -p; GitHub Actions and other CI |
| SDKs | Claude Agent SDK for Python and TypeScript | TypeScript and Python SDKs, Cloud Agents API, Agent SDK |
Context gathering#
Claude Code finds context by searching the repository with its own tools and following references, much as a developer would with grep. MCP tool definitions are deferred by default, so only tool names consume context until a tool is used, and /context shows what fills the window. Cursor's agent uses Instant Grep, a regex engine whose index stays on the developer's machine; Cursor's docs state that it does not store embeddings of the codebase for search. Both hand broad searches to an Explore subagent so raw output stays out of the main conversation (Cursor's runs on a faster model by default), the trade-off context engineering is about.
Instruction files: CLAUDE.md, AGENTS.md, and Cursor rules#
Claude Code reads CLAUDE.md files from the project, the user's home directory, and an organization-managed location, plus path-scoped rules in .claude/rules/. Since version 2.1.277, Claude Code reads AGENTS.md as the project instructions when no CLAUDE.md or CLAUDE.local.md exists in the working directory or above it. AGENTS.md is the cross-tool format that Cursor, Codex, and other agents also read.
Cursor's project rules live in .cursor/rules/ as .mdc files whose frontmatter applies a rule always, by file glob, by the agent's judgment, or by @-mention. Cursor also merges nested AGENTS.md files, and its rules help page says it reads CLAUDE.md the same way and always applies it.
That overlap makes one shared file practical. Keep common instructions in AGENTS.md and add a CLAUDE.md that imports it, so Claude Code loads it even in sessions that can't read AGENTS.md directly:
@AGENTS.md
## Claude Code only
- Run the test suite with `npm test` before proposing a commit.Anthropic's docs also note that a personal CLAUDE.local.md counts as a CLAUDE.md, so adding one to a repository that relies on AGENTS.md alone stops Claude Code from reading AGENTS.md for that developer.
Subagents, skills, hooks, and plugins#
A subagent runs a side task in its own context window and returns a summary. Claude Code adds agent teams for coordinating several full sessions; Cursor adds cloud subagents that run on their own VM and branch. Cursor reads custom subagents from .claude/agents/ and .codex/agents/ as well as .cursor/agents/, with .cursor/ winning name collisions.
Skills follow the open Agent Skills format in both tools: a folder with a SKILL.md file that the agent loads when relevant or a user invokes with a slash command. Cursor loads skills from .claude/skills/, and its third-party hooks support runs hooks from Claude Code's .claude/settings.json files, treating exit code 2 as a block as Claude Code does. Anthropic's docs describe no matching support for .cursor/ directories, so assets meant for both tools belong in Claude Code's locations.
Both connect to MCP servers for issue trackers, databases, and internal APIs, and both package skills, subagents, hooks, and servers into plugins distributed through marketplaces. The Model Context Protocol explainer covers how those servers work.
Checkpoints, approvals, and sandboxing#
Claude Code saves a checkpoint before each prompt and keeps snapshots for the 100 most recent in a session; /rewind, or Escape twice, restores code, conversation, or both. Files changed only by Bash commands are not tracked, so Git remains the real undo. Cursor creates checkpoints before significant changes, stores them locally and separately from Git, and restores files without removing messages.
Both added a classifier model that reviews risky actions in place of a person. Claude Code's auto mode became the starting permission mode for interactive terminal and VS Code sessions in version 2.1.283, while its Bash sandbox for file writes and network access is off by default. Cursor recommends Auto-review, which runs allowlisted calls, sandboxes other shell commands where it can, and sends the rest to a classifier on Cursor's backend; Cursor's docs say Auto-review is not a security boundary. Neither classifier replaces human-in-the-loop review of the final diff before merge.
Code review#
Claude Code's Code Review is a GitHub app for Team and Enterprise plans. Several agents read each change in context, a verification step filters false positives, and findings arrive as inline comments ranked by severity, without approving or blocking the pull request. It is billed on token usage through usage credits, separately from plan usage, and is unavailable with zero data retention. Any plan can run /code-review locally.
Cursor's Bugbot reviews pull requests on GitHub, GitLab, Bitbucket, Azure DevOps, and Origin, reads rules from .cursor/BUGBOT.md, and can push fixes with Autofix. Cursor moved Bugbot to usage-based billing in 2026, and its help center put the average run at $1.00 to $1.50. Agent Review runs a similar pass on local changes.
Headless, CI, and SDKs#
Both run non-interactively with a -p flag for scripts and CI:
# Claude Code: one run with an explicit tool allowlist
claude -p "Find and fix the bug in auth.py" --allowedTools "Read,Edit,Bash"
# Cursor CLI: print mode rejects actions that would need approval unless --force is set
agent -p "review these changes for security issues" --output-format textAnthropic recommends --bare for scripted runs, which skips auto-discovery of hooks, skills, MCP servers, and CLAUDE.md; without it, a -p run executes a repository's project hooks even in an untrusted folder. Claude Code's GitHub Action answers @claude mentions, and Anthropic documents GitLab CI/CD. Cursor documents its CLI in GitHub Actions with a CURSOR_API_KEY secret.
The Claude Agent SDK exposes Claude Code's tools, agent loop, hooks, and permissions as Python and TypeScript libraries. Cursor's TypeScript and Python SDKs run its agent locally or in a cloud VM, with inference on Cursor's hosted models either way, and its Agent SDK, generally available since October 20, 2026, deploys agents defined as code.
Claude Code and Cursor pricing as of November 16, 2026#
Both sell subscriptions with usage limits, but the units differ. Claude plans metered Claude usage in five-hour and weekly windows shared with the Claude chat apps. Cursor's current plans metered a weekly pool for Cursor's own models and billed third-party models, including Claude, per token from prepaid credit at API list price.
On November 16, 2026, Anthropic's pricing page listed these plans with Claude Code:
| Plan | Price | Usage |
|---|---|---|
| Pro | $20 a month, or $17 a month billed annually | Base usage, reset every 5 hours and every week |
| Max | $100 a month | 5x Pro |
| Ultra | $200 a month | 10x Pro per week; no 5-hour limit for verified accounts |
| Team, Standard seat | $20 per seat a month annually, $25 monthly | More than Pro |
| Team, Premium seat | $100 per seat a month annually, $125 monthly | 5x a Standard seat |
| Enterprise | $20 per seat a month annually, 20-seat minimum | Usage billed at API rates |
The free Claude plan did not include Claude Code. Anthropic said how fast a user reaches a limit depends on the model, the conversation length, and file sizes. Team seat allowances reset on a rolling five-hour window and a weekly window, shared with Claude chat and Cowork, and usage credits let paid users continue past them. Claude Code also runs on an Anthropic Console account billed per token, with Opus 5.5 listed at $4 and $20 per million input and output tokens and Sonnet 5.5 at $2 and $10.
On the same day, Cursor's pricing page listed these plans under The X Subscription:
| Plan | Price | Usage and access |
|---|---|---|
| Free | $0 | Limited agent and chat requests |
| Plus | $8 a month | Base pool; Cursor models only; no cloud agents or Bugbot |
| Premium | $30 a month | 4x Plus; cloud agents, Bugbot, Origin, third-party models |
| Super | $100 a month | 15x Plus; priority access at peak times |
| Ultra | $200 a month | 50x Plus; earliest access to new models |
| Teams | $30, $100, or $200 per seat a month | Matches the seat's tier; SSO and analytics |
| Enterprise | Custom | Pooled usage, SCIM, audit logs |
Cursor's pool refreshed weekly and was shared with Grok's chat, image, voice, and Bot products, with different models drawing at different rates. Third-party models sat outside the pool. From Premium up, they were billed at the provider's API list price from the first request, paid from prepaid credit with an optional monthly cap. Teams and Enterprise added a Cursor Token Rate of $0.25 per million tokens on eligible third-party models, including requests made with the team's own key. Cursor's older Pro, Pro+, Ultra, and Teams plans continued for existing customers.
In practice, a developer who wants Claude all day gets that usage inside a Claude subscription's limits, while Cursor's current plans charge Claude usage per token on top of the subscription. A developer happy with Grok or Composer gets them inside Cursor's pool, and on Cursor's Premium, Super, and Ultra plans a personal Anthropic API key moves Claude usage to Anthropic's bill.
Enterprise and governance#
Both products offer SSO, SCIM, audit logs, admin policy, usage analytics, and OpenTelemetry export on their business tiers. The structural difference is the data path. Claude Code sends requests from the developer's machine to a provider or gateway the organization chooses, while every Cursor model request passes through Cursor's backend first.
| Control | Claude Code | Cursor |
|---|---|---|
| SSO | Team and Enterprise | SAML or OIDC on Teams and Enterprise |
| SCIM and audit logs | Enterprise | Enterprise |
| Central policy | Managed and server-managed settings, MCP controls | Team settings, MDM policies, managed hooks, MCP allowlist |
| Model access | availableModels allowlist in managed settings |
Model access control and BYOK limits on Enterprise |
| Training and retention | Commercial terms: no training, 30-day standard retention, zero data retention for qualified Enterprise accounts | Privacy Mode: no training under zero data retention agreements; on by default for Enterprise |
| Usage analytics | Team, Enterprise, and Console dashboards; OpenTelemetry | Teams and Enterprise dashboard; Analytics API and OpenTelemetry on Enterprise |
| Data residency | US-only inference on the Anthropic API at 1.1x token pricing | US-only residency on Enterprise at a 10% model price uplift |
Data handling. Anthropic's data usage page separates consumer accounts (Free, Pro, Max, and Ultra), whose data trains future models when the user leaves that setting on, from commercial accounts (Team, Enterprise, API, and cloud providers), whose code and prompts Anthropic does not train on. Claude Code also keeps transcripts locally under ~/.claude/projects/ for 30 days by default. Cursor's Privacy Mode guarantees code is not used for training; its cloud agents store encrypted repository copies that are deleted when the agent finishes, and its zero data retention terms do not cover requests made with your own API key.
Gateways. Running Claude Code through an AI gateway is a documented deployment pattern. Anthropic's LLM gateway guide has administrators issue each developer a gateway credential and distribute the address through managed settings:
export ANTHROPIC_BASE_URL=https://llm-gateway.example.com
export ANTHROPIC_AUTH_TOKEN=sk-gateway-keySetting allowedProviders to ["customEndpoint"] in managed settings makes that gateway the only destination a managed machine accepts. Anthropic also ships a self-hosted gateway, Claude apps gateway, inside the claude binary, and third-party gateways work too; Bifrost, for example, documents Claude Code, Cursor, and other agents in its CLI agents guide. Cursor's guidance runs the other way. Its network configuration docs warn that custom gateways can add latency, rate limiting, and compatibility problems and recommend hooks instead. Its API key page says every request routes through Cursor's servers for final prompt building, even with your own key. Teams that must log every prompt in their own systems can do so at the gateway for Claude Code, and with hooks and OpenTelemetry export for Cursor.
Codex vs Cursor and Claude Code#
OpenAI's Codex sits closer to Claude Code than to Cursor: it is a model vendor's agent, sold with that vendor's chat subscription and centered on its own models, with a CLI, IDE extension, desktop app, and cloud runs. Compared with Claude Code, it differs in its open-source CLI and OpenAI models; compared with Cursor, it is not an editor.
The Codex CLI is open source under the Apache 2.0 license, and codex exec runs it in scripts and CI. OpenAI's IDE extension installs in VS Code, Cursor, and Windsurf; Codex also runs in the ChatGPT desktop app, on the web, and as Codex Cloud; and it reads AGENTS.md. OpenAI's Codex pricing page on November 16, 2026 included Codex in ChatGPT plans from Free and Go ($8 a month) through Plus ($20), Pro (from $100), and Business ($20 per user a month, billed annually). The Codex vs Cursor decision follows the same logic: Codex for OpenAI's models in a terminal-first agent, Cursor for an editor with several vendors, or Codex's extension inside Cursor.
Which should you choose?#
Choose by where your developers work and which models and controls your organization needs. Terminal-first teams that standardize on Claude, or must route traffic through their own gateway or cloud account, fit Claude Code. Editor-first teams that want inline completion and a choice of model vendors fit Cursor. Many teams run both.
| If your team | Lean toward | Why |
|---|---|---|
| Works mainly in terminals, SSH sessions, and scripts | Claude Code | The CLI is its most complete surface |
| Wants an editor with inline completion and visual diffs | Cursor | Tab, the editor, and the Agents Window |
| Standardizes on Claude or buys it through Bedrock, Google Cloud, or Foundry | Claude Code | Direct support for those providers |
| Wants several model vendors or Cursor's own models | Cursor | Multi-vendor catalog and Auto mode |
| Must keep all model traffic on an approved gateway | Claude Code | ANTHROPIC_BASE_URL plus allowedProviders |
| Reviews pull requests on GitLab, Bitbucket, or Azure DevOps | Cursor | Bugbot supports them; Code Review is GitHub only |
Before committing, run the same 10 to 30 closed issues through both tools with the same instruction file, as the AI coding agents comparison describes, and check the AI coding tools directory for other agents worth adding to the trial.
Using Claude Code and Cursor together#
The common split is Cursor for editing and review and Claude Code in the terminal and CI. Because Cursor reads Claude Code's instruction files, skills, subagents, and hooks, one set of project configuration serves both, and the Claude Code extension runs inside Cursor itself.
- Install Claude Code inside Cursor. Use the extension's Cursor install link, or run
claudein Cursor's terminal. It bills your Claude plan or Console account, not Cursor's pool. - Share one instruction file. Keep shared rules in
AGENTS.mdwith aCLAUDE.mdthat imports it, and put Cursor-only, glob-scoped rules in.cursor/rules/. - Store shared assets in Claude Code's folders. Skills in
.claude/skills/, subagents in.claude/agents/, and hooks in.claude/settings.jsonload in both, as long as Cursor's third-party import setting stays on, its default. - Divide the loop. Use Cursor for exploratory edits, Tab, and diff review; use Claude Code for long terminal tasks,
claude -pjobs, and the GitHub Action. - Watch both bills. Claude Code draws on the Claude plan, while Claude models picked inside Cursor bill per token on Cursor's current plans.
For more comparisons of terminal, IDE, and cloud agents as they change, follow the AI coding topic hub.
Frequently asked questions
Is Claude Code better than Cursor?
- Neither is better for every team. Claude Code suits developers who work in a terminal, standardize on Claude models, or need to route traffic through their own gateway or cloud provider. Cursor suits developers who want an editor with inline completion, visual diffs, and a choice of model vendors. Both can run the same Claude models, so a short trial on your own repositories is the most reliable test.
Can you use Claude Code inside Cursor?
- Yes. Anthropic's documentation provides an install link for the Claude Code extension in Cursor, because Cursor is based on the VS Code codebase and accepts VS Code extensions. You can also run the claude command in Cursor's integrated terminal. The extension and the CLI use your Claude subscription or API account, not Cursor's usage pool.
Does Cursor use Claude models?
- Yes. Cursor's model catalog listed Claude models, including Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.2, on November 16, 2026. On Cursor's current plans, third-party models like Claude were available from the Premium plan up, billed at the model's API list price from prepaid credit. The Plus plan included only Cursor's own models. Premium, Super, and Ultra subscribers could also add their own Anthropic API key.
Is Claude Code free?
- No. Anthropic's pricing page on November 16, 2026 did not include Claude Code in the free Claude plan. Claude Code came with the Pro plan at $20 a month billed monthly, and with Max, Ultra, Team, and Enterprise plans. Developers can also use Claude Code with an Anthropic Console account and pay per token at API rates instead of a subscription.
Can Claude Code and Cursor share the same project instructions?
- Yes. Cursor reads AGENTS.md and CLAUDE.md files, and Claude Code reads AGENTS.md when a repository has no CLAUDE.md. A common setup keeps the shared instructions in AGENTS.md and adds a CLAUDE.md that imports it with an @AGENTS.md line. Cursor also loads skills from .claude/skills and subagents from .claude/agents, so those can live in one place.
Who owns Cursor?
- Cursor is built by Anysphere, which SpaceX acquired. Cursor's blog announced on August 14, 2026 that the acquisition had closed, completing a process that began with a model-training partnership in April 2026. By November 2026, Cursor listed Grok models among its own models and sold its plans as part of The X Subscription, which combines Cursor, Grok, and X.
Terms in this guide
Cite this guide
The AI Engineering Editors. "Claude Code vs Cursor: how the two coding agents differ." The AI Engineering, updated Nov 16, 2026. https://theaiengineering.com/articles/claude-code-vs-cursor
- Title
- Claude Code vs Cursor: how the two coding agents differ
- Author
- The AI Engineering Editors
- Publisher
- The AI Engineering
- Published
- Updated