Comparison ยท AI coding agents

Cursor vs Windsurf vs Claude Code

Three different shapes of coding agent: an editor with agents built in, an agent command centre that grew out of an editor, and a terminal-first agent that now runs on the web and desktop too. Pricing is close; workflow and governance are not.

Short answer. Pick Cursor if you want a polished editor with agent, cloud and CLI surfaces and you are comfortable paying for model usage on top of a seat. Pick Devin Desktop โ€” the product formerly called Windsurf โ€” if you want one command centre for several agents and cloud sessions. Pick Claude Code if you want a terminal-first agent with strong repo-wide context, hooks and subagents, and you are already on a Claude plan. Seat prices start at $8, $20 and $20 per month respectively, before model overages.

Side by side

DimensionCursorWindsurf / Devin DesktopClaude Code
Form factorIDE, CLI, cloud agentsDesktop command centre plus IDETerminal, IDE, desktop, web
Free tierLimited agent and chat requestsLight agent quotaNo Claude Code on Free
Entry paid$8 Plus$20 Pro$20 Pro ($17 annual)
Top individual$200 Ultra$200 Max$200 Ultra
Team price$30+/user$80/month + $40/full seatTeam and Enterprise via sales
Model choiceGrok included; third-party at list priceFrontier and open modelsClaude models; supported third-party providers
ContextModel-dependentModel-dependentUp to 1M, varies by model
Best forEditor-centric teamsMulti-agent oversightTerminal and CI-heavy teams

Sources: Cursor pricing, Devin Desktop pricing (Windsurf's current home), Claude pricing and Claude Code docs, checked 20 November 2026. Seat prices exclude taxes.

What each one is now

Cursor is an editor with agent modes, a CLI and cloud agents. Its pricing page now sits under xAI and bundles Grok models into the seat; third-party models are available at API list price. It is the most integrated option for editing, chat and delegated tasks.

Windsurf is now Devin Desktop. Cognition renamed the product and repositioned it as a command centre for local and cloud agents, with the full IDE still available. Existing Windsurf plans and settings carry over. If you are evaluating "Windsurf" from an old comparison, you are really evaluating Devin Desktop's agent management and its Pro, Max and Teams plans.

Claude Code started as a terminal agent and now runs in the terminal, VS Code, JetBrains, a desktop app and the browser. It reads the repo, edits files, runs commands and integrates with git, MCP servers and CI. It is included on Claude Pro and above, subject to five-hour and weekly usage limits.

Agent modes

Cursor's agent can plan, edit multiple files, run commands and hand work to a cloud agent. Cloud agents are useful for queued tasks, but they need a clean repository, tests and a review process to be safe.

Devin Desktop's differentiator is oversight: a board of sessions, local and cloud handoff, and support for multiple agents through the Agent Client Protocol. It is designed for a developer running several agents at once, not just one chat pane. That is valuable for large refactors and parallel bug fixes; it also multiplies the number of diffs a reviewer must handle.

Claude Code is the most composable. A lead agent can spawn subagents, hooks can run formatters or linters before and after edits, and CLAUDE.md or an existing AGENTS.md carries project instructions. The CLI pipes into Unix tools, which makes it a natural fit for CI and scripted maintenance.

Model choice and context

Cursor bundles Grok usage and lets you buy third-party models at API list price. Devin Desktop supports frontier and open models and exposes a "Fusion" mode that pairs a lead model with a cheaper sidekick. Claude Code is Claude-first, with Opus, Sonnet and Haiku tiers and up to 1M tokens of context depending on model and plan.

Context is not a free win. A 1M-token window raises cost and can dilute attention. The better pattern is retrieval plus a focused context: the files that matter, the failing test, the interface contract and the last review comments. Test which tool finds the right files in your monorepo.

Pricing and TCO per developer

Seat price is only the first line. The second line is model usage: included quotas, pass-through API pricing or subscription limits. The table below models a ten-developer team on the entry team plan, excluding taxes and any extra usage.

PlanSeat cost10 developersModel usage
Cursor Teams$30/user/month$300/monthGrok included; third-party at API list price
Devin Desktop Teams$80/month + $40/full seat$480/monthQuotas included; extra usage at API pricing
Claude Code Pro$20/user/month$200/monthUsage limits every 5 hours and weekly
Claude Code Max$100/user/month$1,000/month5x Pro usage; higher weekly limit

TCO also includes review time, CI minutes and the engineering owner for permissions. A $20 seat that produces unreviewable diffs costs more than a $100 seat that fits the workflow.

Privacy and administration

All three vendors publish enterprise controls, but the details differ. Cursor offers a privacy mode that prevents code from being used for training by Cursor or its model providers, plus team-wide settings and SSO on higher tiers. Devin Desktop offers SSO, cost controls, VPC deployment and enterprise administration. Claude Code's individual plans lack SSO and central billing; those controls sit in Claude's Team and Enterprise offerings.

Ask four questions before rollout: where is code processed, how long are prompts retained, can training be disabled contractually, and which subprocessors see the data? A vendor's trust centre is a starting point, not a substitute for reading the current DPA and security addendum.

When to use the API directly

An IDE agent is optimised for a human in the loop. A custom agent is better for a CI job, a nightly refactor or a review bot that runs without a developer watching, with deterministic permissions and your own logs.

from openai import OpenAI

client = OpenAI(
    api_key="PLUGSKY_KEY",
    base_url="https://api.plugsky.com/v1",
)

resp = client.chat.completions.create(
    model="plugsky-coder",
    messages=[
        {"role": "system", "content": "You are a careful code reviewer. Return JSON with file, line, severity and fix."},
        {"role": "user", "content": "Review this diff for null handling:\n" + diff},
    ],
    response_format={"type": "json_object"},
)
print(resp.choices[0].message.content)

FAQ

Is Windsurf still called Windsurf?

Windsurf is now Devin Desktop. Cognition renamed the IDE after acquiring it, and existing plans, pricing, extensions and settings carry over. The product is still available, but current documentation and billing use the Devin name.

Which is cheapest per developer?

On seat price alone, Claude Code Pro at $20 per month and Devin Desktop Pro at $20 per month are lower than Cursor's $30 Teams seat. Cursor's seat includes Grok usage but bills third-party models at API list price, so the all-in cost depends on how much frontier-model usage each developer consumes.

Which has the largest context window?

Claude Code advertises up to 1M tokens of context, varying by model and plan. Cursor and Devin Desktop inherit the context limit of the model you select, so the answer changes when you switch models. Large context is not free: it increases latency and token cost.

Do these tools work with third-party models?

Cursor and Devin Desktop support multiple model providers, and Claude Code can run against Anthropic models or through supported third-party providers. For a fully custom agent, call an OpenAI-compatible API directly and keep the harness in your own code.

When should I use the API instead of an IDE agent?

Use the API directly for CI jobs, batch refactors, custom review bots and any workflow where you need deterministic permissions, your own logging or a fixed monthly cost. An IDE agent is better for interactive work where a developer reviews each diff.

Reviewed 20 November 2026. Prices and plan features were checked against Cursor pricing, Devin Desktop pricing (Windsurf's current product), Claude pricing and Claude Code docs. Verify before renewal.