Cursor vs Antigravity vs Claude Code vs Windsurf

Cursor, Google Antigravity, Claude Code, and Windsurf can all edit code with an agent, but they are not four versions of the same product. Cursor and Windsurf put an AI agent inside a dedicated code editor. Antigravity offers an IDE alongside a separate command center and CLI. Claude Code is an agent that can run in a terminal, an existing IDE, Anthropic's desktop app, or the web.

That surface difference should drive the decision. Choose Cursor or Windsurf if you want the editor to be the center of daily work. Consider Antigravity if browser operation, artifacts, and parallel local agents belong in the same Google product family. Start with Claude Code if you want an agent to follow your existing terminal and editor setup instead of replacing it.

Vendor documentation for this guide was reviewed on July 17, 2026. No HUMAI benchmark, generated-code score, or private production repository is presented as evidence. Plan names, usage pools, and model access can change, so confirm them on the linked vendor pages before buying.

Start with where the work lives

The compact table below compares product shape, not model quality. A model available in two tools can behave differently because each product supplies its own context, tools, permission system, and execution environment.

The table is intentionally wider than a phone screen. HUMAI's theme keeps tables inside their own horizontal scroller on narrow displays.

Cursor keeps editing and delegation in one account

Cursor's local product is a code editor with Tab completion, inline editing, and an Agent panel. Its current documentation describes Agent as able to search a codebase, edit files, and run terminal commands, while the review interface exposes the resulting diff. Project instructions can live in version-controlled Cursor Rules, with user rules available for preferences that should follow one developer.

The product now extends well beyond a single desktop session. Cursor's January 2026 CLI update documents local planning, read-only Ask mode, and handoff to cloud agents. A developer can begin with repository context locally, send a task to the cloud, and later resume it on the web, mobile, or in the editor.

That handoff changes the security question. Cursor's June 2026 cloud-agent update says a cloud subagent runs in its own virtual machine and branch and can later be pulled back to the local computer. A separate self-hosted route keeps code and tool execution inside the customer's network. Review repository access, environment setup, secrets, network policy, and the return to human review for the exact deployment.

For confidential code, read Cursor's security description and Data Use and Privacy overview before enabling indexing or remote work. The security page describes infrastructure controls, Privacy Mode, and where subprocessor information is maintained. The data-use page says requests still pass through Cursor's backend and explains that codebase chunks are uploaded to compute embeddings, plaintext ceases to exist after the request, and embeddings plus metadata may be stored. A team should test its actual settings and contract rather than infer data handling from the selected model name.

The live Cursor pricing page separates individual, team, and enterprise plans. Plans include a quantity of model usage, and on-demand usage can continue after that allowance. Team and enterprise tiers add administration and control features. Price a representative month that includes foreground Agent work, cloud runs, review, and any code-review product your team plans to use.

Antigravity is a product family, not one IDE window

Google's current documentation separates four Antigravity surfaces: a standalone Antigravity 2.0 command center, Antigravity IDE, Antigravity CLI, and an SDK. The IDE contains an editor, Tab completion, agents, terminal access, browser operation, and visual artifacts. The separate Antigravity 2.0 application is built to launch, monitor, and orchestrate agents independently of the IDE.

This distinction matters when a comparison calls Antigravity an IDE and stops there. A developer can still work directly in the IDE, but the wider product is designed around managing agents across projects. Google defines a Project as one or more folders with its own settings and permissions. A conversation can work in the active checkout or in a new Git worktree, which is useful when parallel agents should not write into the same branch.

Antigravity also gives the browser a formal role. Its browser documentation says a browser subagent can read and operate a separate Chrome profile, capture screenshots, and save action recordings as artifacts. That can fit UI verification and web-based development work, but it also expands the permission surface. Browser tools can be disabled, and URL allowlists and denylists should be treated as part of project configuration.

The most useful differentiator is not a claim of greater autonomy. It is the visibility and control model. Antigravity Artifacts include plans, code diffs, diagrams, images, and browser recordings. Its IDE settings expose terminal review, access outside project folders, browser JavaScript review, artifact approval, and terminal sandboxing. Strict Mode forces review-oriented settings, confines file access, and enables a sandbox with network access denied.

Google's current plan page lists individual access, Google AI Pro, Google AI Ultra, and an organization route through Google Cloud. The plans use rate limits or credit pools rather than a promise of unlimited complex agent work. Confirm eligible accounts, current models, regional availability, and the organization terms that will govern code before rollout.

Claude Code does not require a new editor

Anthropic describes Claude Code as an agentic coding tool that reads a codebase, edits files, runs commands, and connects to development tools. The full CLI runs in a terminal. Anthropic also provides a VS Code extension, JetBrains integration, a desktop interface, and a browser surface. That makes Claude Code a broader product than a terminal-only utility, but it still does not supply the underlying code editor in the way Cursor or Windsurf does.

The distinction gives teams a migration option. The VS Code integration provides inline diffs, file references, plan review, conversation history, worktree support, and MCP connections. Anthropic documents installation in VS Code, Cursor, and other compatible forks. A team can therefore evaluate Claude Code inside an existing editor before deciding whether any editor replacement is justified.

Claude Code's permission model is especially legible. Anthropic's security guide says it begins with strict read-only permissions and asks before file edits, commands, and broader access. It documents a working-directory boundary, sandbox options, network-command approval, trust checks for new codebases and MCP servers, and the user's responsibility to review actions. These controls do not make generated changes correct, but they give a security review concrete settings to inspect.

Account type affects both cost and data terms. Anthropic's plan page says Claude Code is included in paid Claude plans and shares usage with other Claude surfaces. API or Console use follows token billing instead. The cost guide recommends tracking actual usage and notes that model choice, context, parallel sessions, and automation change spend. Anthropic's data-usage page also separates consumer training preferences from commercial Team, Enterprise, API, and supported third-party-provider terms.

Windsurf 2.0 connects its editor to Devin

Windsurf's current first-party page still presents Windsurf 2.0 as an IDE. It combines a local editor, Cascade, Tab completion, an Agent Command Center, and a handoff from local work to Devin, Cognition's cloud agent. The command center organizes local and cloud agents, sessions, pull requests, files, and project context.

Cascade remains the editor-side agent. The current Cascade documentation describes Code and Chat modes, terminal and MCP tools, planning, queued messages, checkpoints, reverts, and simultaneous sessions. Code mode can change a repository; Chat mode is oriented toward questions and proposed code. When simultaneous agents might touch the same files, the documentation recommends worktrees to prevent collisions.

The cloud handoff is the larger boundary. Windsurf says Devin runs in its own virtual machine with a desktop, browser, and computer-use tools, and can continue after the local laptop closes. Before enabling that path, verify repository permissions, secrets, network access, review ownership, and the event that returns work to a human. A local plan becoming a cloud run is an execution change, not merely a user-interface shortcut.

There is also a documentation transition to notice. Older Windsurf documentation URLs now resolve into Devin Docs, while the public Windsurf pricing URL resolves to the wider Devin pricing page. The dedicated Windsurf account page still lists a free tier, a paid Pro route, model access, and cloud agents. For an individual trial this is manageable. For procurement, confirm the current product name on the order, which quotas cover Cascade and Devin, and which Cognition security terms apply.

Permissions matter more than autonomy labels

All four products can move from explanation to action. The practical question is what the agent can read, modify, execute, and send over the network at each stage. Marketing words such as agentic or autonomous do not answer that.

• Repository scope: record which folders, repositories, submodules, and worktrees are visible.
• Command scope: distinguish read-only inspection, approved local commands, allowlisted commands, and unattended execution.
• Network scope: identify browser access, package downloads, web search, MCP servers, and outbound connections from remote machines.
• Secret scope: supply task-specific credentials with the smallest useful permissions and a clear revocation path.
• Review scope: decide who approves a plan, who reads the diff, who checks generated migrations, and who can merge.

A product can be suitable for a public demo repository and unsuitable for a regulated codebase under the same account. Run the permission review for the intended surface. Cursor foreground and background agents differ. Antigravity IDE, browser, and Strict Mode differ. Claude Code consumer and commercial routes differ. Windsurf local Cascade and Devin cloud execution differ.

Treat pricing as a workload equation

A monthly sticker price does not reveal the cost of an agent workflow. Cursor combines plan allowances with possible on-demand usage. Antigravity uses access tiers, rate limits, or credits. Claude Code can draw from a Claude subscription pool or incur token charges through a Console or cloud route. Windsurf and Devin plans include quotas whose consumption depends on the model and job.

Use one representative workload to estimate cost: a bounded issue, the files required to solve it, the commands allowed, and the review steps required before merge. Record the number of agent runs, retries, context resets, cloud handoffs, and human review minutes. Do not convert vendor request allowances into a universal task count because tasks, models, and context lengths differ.

For teams, add costs outside the prompt box. Include identity management, policy enforcement, audit access, repository connectors, environment maintenance, secret rotation, and the engineering time needed to review generated changes. A cheaper plan can be the more expensive operating choice if it leaves controls or handoffs to manual work.

A seven-day trial that produces evidence

Run one tool at a time on the same disposable repository. The goal is not to crown a winner from a synthetic benchmark. It is to discover whether the product's surface and controls fit your work.

1. Choose three existing issues: one localized fix, one cross-file change, and one task that requires a browser or external documentation.
2. Write acceptance checks before prompting. Include tests, linting, file boundaries, forbidden changes, and the expected review artifact.
3. Create a clean branch or worktree for every run. Do not allow agents to share an uncommitted working directory.
4. Start with minimal permissions. Log every additional folder, command, network destination, secret, or repository permission requested.
5. Save the plan, diff, terminal transcript, test output, usage record, and reviewer comments. A final passing test without the path to it is insufficient evidence.
6. Have a developer who did not prompt the agent review the change. Count unresolved questions and corrections, not generated lines.
7. Remove the tool's repository access and secrets at the end, then verify that the offboarding process actually works.

The trial should end with a short record for each product: acceptable surfaces, prohibited repositories, required settings, spending observations, and stop conditions. That record is more useful than a score assembled from vendor feature lists.

Make the choice at the handoff point

If developers want to stay inside one dedicated editor and sometimes delegate work to remote agents, evaluate Cursor first. If a Google-centered team needs an IDE plus a separate command center, browser subagents, artifacts, and explicit project security presets, evaluate Antigravity. If the team wants to keep its terminal and editor while adding a permission-oriented agent, evaluate Claude Code. If local editing should hand work directly to a Cognition cloud agent, evaluate Windsurf 2.0 and document where Cascade ends and Devin begins.

Do not buy all four to avoid choosing. Pick the handoff that matters, run the controlled trial, and keep the tool only if its permissions, evidence, and billing can be explained before the first private repository is connected.