IDE and cloud coding agents compared
Cursor, VS Code agent mode, Google Antigravity, Devin and Kiro compared on local versus cloud execution, background tasks, pull-request workflow, model choice, MCP and skills, pricing, and enterprise and privacy controls.
Comparison
| Entry | Where it runs | Background and async work | PR workflow | Model choice | MCP and skills | Pricing | Enterprise and privacy |
|---|---|---|---|---|---|---|---|
| Cursor | Local agent in the Cursor editor and CLI; Cloud Agents in isolated Cursor VMs; or Self-Hosted Machines that run the tools on your own workers while Cursor runs the agent loop | Run many Cloud Agents in parallel without your machine online. Start them from desktop, web, iOS, Slack, GitHub or Bitbucket comments, Linear or the API; Automations | Cloud agents clone from GitHub, GitLab, Azure DevOps or Bitbucket, work on a branch and open pull requests, including multi-repo; Bugbot reviews PRs | Cursor models (Composer, Grok) plus OpenAI, Anthropic and Google models; Cloud Agents use a curated subset | MCP over HTTP and stdio with OAuth, locally and in Cloud Agents; Agent Skills from .agents/.cursor/.claude/.codex skills folders | Free; Plus $8/mo; Premium $30; Super $100; Ultra $200; Teams $30–$200 per seat; Enterprise custom. Third-party models billed at API list price | Privacy Mode (no training by Cursor or providers, on by default for Enterprise), team enforcement, allowed team IDs, network allowlists, private connectivity, enterprise hooks |
| VS Code Agents (agent mode) | In the editor (Local or Copilot harness), in worktrees, or the Cloud target on provider-managed infrastructure; Claude and Codex harnesses also available | Agents window manages multiple sessions; Copilot CLI and the Codex harness support background work; cloud sessions keep running remotely | The Cloud target works against a GitHub repository and returns a pull request; local sessions leave diffs for you to review and commit | Models from your Copilot plan, your own API key (BYOK) or a supported local model; the harness and model are separate choices | MCP servers; Agent Skills from .github/.claude/.agents skills and ~/.copilot/skills | The editor is free; agent usage depends on your GitHub Copilot plan or BYOK provider (plan prices not covered here) | Enterprise AI policies can restrict agents, models and tools on managed devices; permissions and approvals; terminal sandboxing options |
| Google Antigravity | Locally in the Antigravity App, IDE, CLI and IDE extensions (VS Code, JetBrains, Visual Studio, Zed, Xcode), with a terminal sandbox | The App manages background agents and subagents, with /schedule and /goal; Remote Control lets you drive desktop and CLI sessions from a browser | Not documented on the pages checked; review happens through artifacts (plans, walkthroughs, screenshots) | Gemini models plus Claude Sonnet and Opus and GPT-OSS 120B, with low, medium and high thinking levels; availability depends on plan | MCP in the App, CLI and IDE; Agent Skills (SKILL.md folders), rules, plugins, hooks and a marketplace | Individual $0 with basic weekly limits; higher limits with Google AI Pro or Ultra; business via Google Cloud (Gemini Enterprise from $30 per seat per month, or pay-as-you-go) | Google Cloud Terms of Service on business plans, spend caps, centralised admin controls; permissions and sandbox |
| Devin CLI | Devin cloud sessions on Cognition VMs; Devin CLI runs locally and hands off to cloud; Outposts run session commands on your own infrastructure while inference stays in Devin's cloud | Fully asynchronous sessions started from the web app, Slack, Teams, Linear, Jira, PagerDuty and others; parallel managed sessions; playbooks | Creates PRs and MRs on GitHub, GitLab and Bitbucket, responds to review comments, stacked PRs, PR templates; Devin Review | In the CLI, a model picker plus Adaptive routing and Fusion pairings; a ChatGPT plan can be linked to bill GPT usage | MCP marketplace plus custom stdio, SSE and HTTP servers; Skills from .agents/skills (recommended) or .devin/skills | Free; Pro $20/mo; Max $200/mo; Teams from $80/mo (up to 200 users); Enterprise via contract with ACU limits | SOC 2 Type II; paid plans can opt out of training, Enterprise is never trained on without consent; security profiles restrict network, MCP and git; self-hosted SCM |
| Kiro | Kiro IDE and CLI locally, or cloud sessions in a managed sandbox that the IDE, CLI, web and mobile can attach to | Up to 10 concurrent cloud sessions that keep running when clients disconnect; autonomous mode; scheduled automations on web | Kiro Web opens pull requests or merge requests; autonomous mode runs from plan to PR | Auto plus Claude, GPT and open-weight models; the free tier is limited to Claude Sonnet 4.5 and open-weight models | MCP (with enterprise MCP governance); skills in .kiro/skills and ~/.kiro/skills | Free tier with 50 credits; paid Pro, Pro+, Pro Max and Power tiers on a credit model; cloud sessions add no compute charge | Enterprise governance for models, MCP, permissions and sign-in, IdP integration, customer-managed KMS keys; individual content may be used for service improvement unless you opt out |
Verdict: For the most ways to run in the cloud, from managed VMs to your own machines, use Cursor. To stay inside your editor and Copilot plan, use VS Code agent mode; it can also run Claude and Codex. Antigravity is local-first, with a free tier and strong artifact review, but no documented PR flow. Devin is the most fully asynchronous cloud teammate, with deep ticket-to-PR integrations. Kiro bundles cloud sessions into its credit plans. All five support MCP and Agent Skills.
The axis this covers
The terminal coding agents comparison is about CLIs on your machine. This one asks three different questions:
- Where does the agent run? In your editor on your laptop, in a vendor VM, or on a machine you own.
- Does it keep working when you leave? Background and cloud sessions survive a closed laptop, and you check in from web, mobile or chat.
- How does work come back? As a diff in your editor, or as a pull request on GitHub, GitLab or Bitbucket.
Execution models
- Editor-local: VS Code agent mode, the Antigravity App and IDE, Kiro IDE, and Cursor's local agent. They have fast feedback and full access to your environment, but they stop when the machine sleeps. Antigravity's Remote Control lets you drive local sessions from a browser.
- Vendor cloud: Cursor Cloud Agents, Devin sessions, Kiro cloud sessions, and VS Code's Cloud target. You get parallel tasks, PRs as output, and triggers from Slack, issue trackers or an API. Your environment setup decides whether the agent can actually test its work.
- Hybrid, on your hardware: Cursor Self-Hosted Machines and Devin Outposts keep execution (checkout, builds, credentials) on your machines while the agent loop runs in the vendor cloud. Use them for private networks or special hardware.
Notes
- Skills are portable. Every product here reads
SKILL.mdskills (see Agent Skills). Cursor, VS Code and Devin all read.agents/skills/; Kiro uses.kiro/skills/. Installers and packs are covered in Agent skill packs, skills CLIs and registries. - Privacy defaults differ. Cursor turns Privacy Mode on by default only for Enterprise. Devin and Kiro may use individual users' content unless you opt out (Devin: paid plans can opt out).
- Not covered: GitHub Copilot coding agent, Windsurf and Jules have no entries in the index yet. VS Code's Cloud target is the documented way to reach a GitHub cloud agent from the editor.
Directory entries in this comparison
- Cursor: AI code editor and coding agent with Agent mode, Rules, Skills, MCP servers and a CLI.
- VS Code Agents (agent mode): Agent mode in Visual Studio Code: an agent finds relevant code, edits across files and runs checks in your workspace while you review the changes.
- Google Antigravity: Google's agentic coding platform (app, IDE, CLI, IDE extensions and Python SDK) for orchestrating autonomous subagents, scheduled tasks and code verification.
- Devin CLI: Cognition's local coding agent that runs in your terminal on macOS, Linux, WSL and Windows, with handoff to cloud Devin sessions.
- Kiro: Agentic engineering environment that turns prompts into executable specs and builds across large codebases with parallel agents.
- GitHub Copilot CLI: Brings GitHub Copilot's coding agent to your terminal to read, write and run code where you work.
- DeepWiki MCP: Official DeepWiki MCP server: programmatic access to DeepWiki's public repository documentation and search (Ask Devin).
- Agent Skills Specification: Open format for packaging agent capabilities as folders with a SKILL.md (YAML frontmatter + instructions) plus optional scripts and resources.
- Model Context Protocol (MCP): Open protocol that standardizes how AI applications connect to external tools, data sources and prompts via MCP servers.
Related
- Terminal coding agents compared (Comparison)
- Agent skill packs, skills CLIs and registries (Comparison)
- Code sandboxes for AI agents (Comparison)
- Coding agent starter stack (Stack)
Sources
- Cursor Cloud Agents, accessed
- Cursor Self-Hosted Machines, accessed
- Cursor models and pricing, accessed
- Cursor privacy and data governance, accessed
- Cursor MCP, accessed
- VS Code, Build with AI, accessed
- VS Code agent harnesses, accessed
- VS Code Agent Skills, accessed
- Antigravity pricing, accessed
- Antigravity, Choose a model, accessed
- Antigravity MCP, accessed
- Antigravity Agent skills, accessed
- Antigravity Remote Control, accessed
- Antigravity llms.txt, accessed
- Devin docs index, accessed
- Devin self-serve plans, accessed
- Devin CLI, accessed
- Devin Outposts, accessed
- Devin Skills, accessed
- Security at Cognition, accessed
- Kiro cloud sessions, accessed
- Kiro Web, accessed
- Kiro skills, accessed
- Kiro pricing, accessed
- Kiro data protection, accessed