# Cursor: features, protocols, quickstart

## TL;DR

Cursor is an AI coding editor and CLI from Anysphere. For multi-agent work it offers an Agents Window for parallel agents, isolated Git worktrees, subagents, cloud agents on their own VMs, and Projects, where a coordinator agent delegates to other agents. This entry covers those features, not the editor as a whole. It suits teams running coding agents in parallel.

## Key facts

| Field | Value |
| --- | --- |
| Type | Platform |
| Languages / SDKs | TypeScript, Python |
| License | Proprietary |
| Pricing model | [Hosted, paid](https://cursor.com/pricing) |
| Orchestration pattern | Supervisor |
| Repository | No public repository (closed source) |
| Website | [cursor.com](https://cursor.com) |
| Documentation | [cursor.com](https://cursor.com/docs) |
| Last verified | 2026-09-30 |

## Key features

- The Agents Window is an agent-first interface for running many parallel agents across local, cloud and remote SSH environments, with a diffs view and hand-off between local and cloud. ([source](https://cursor.com/docs/agent/agents-window))
- Worktrees run each agent task in its own Git checkout so several agents can start on one repository without conflicts; `.cursor/worktrees.json` holds setup commands. ([source](https://cursor.com/docs/configuration/worktrees))
- Cloud agents run in isolated VMs with cloned repositories, secrets and startup commands, and can be started from the desktop app, web, iOS, Slack, GitHub, Linear or the API; the docs say any number can run in parallel. ([source](https://cursor.com/docs/cloud-agent))
- Subagents have their own context windows, run in the foreground or background, can be defined in `.cursor/agents/` (Claude and Codex agent directories are also read), and can be launched in parallel or handed to a cloud VM with `/in-cloud`. ([source](https://cursor.com/docs/subagents))
- Projects use a coordinator agent that plans work and delegates it to many agents running on Cloud Agents, with shared context files that sync across the agents' machines. ([source](https://cursor.com/docs/agent/projects))
- TypeScript and Python SDKs wrap local and cloud runtimes behind one Agent interface, and a Cloud Agents HTTP API (public beta) starts and manages cloud runs. ([source](https://cursor.com/docs/sdk/typescript))
- Cursor CLI can run as an Agent Client Protocol server over stdio with `agent acp`, so a custom client can drive it. ([source](https://cursor.com/docs/cli/acp))

## Architecture and orchestration pattern

Pattern: Supervisor.

Cursor is a product with several ways to run more than one agent. On the desktop, the Agents Window lists agents across projects and environments; each agent can work in its own Git worktree, or in a cloud VM that has its own branch. The docs describe the agent loop as instructions, tools and a model, with Cursor tuning the prompts and tools per model. There is no user-defined graph or workflow language.

Two delegation mechanisms exist. Subagents are child agents that the main agent launches for context-heavy or parallel work: each has its own context window, starts clean with only the prompt the parent gives it, and returns a final message. Built-in Explore, Bash and Browser subagents run automatically; custom ones are markdown files with frontmatter. Projects add a coordinator agent that does not write code itself; it creates and manages agents on Cloud Agents and can start a local agent when something must be tested on your machine.

State is kept per agent (conversation, branch, environment build). A Project keeps a set of shared context files that sync to every cloud and local machine its agents use. Cloud agents store code and environment data in Cursor's cloud while they run, and Cursor manages VM provisioning and snapshots.

### Human in the loop

Run Modes control when Cursor interrupts for approval of shell, MCP and Fetch tool calls: Auto-review (allowlist, sandbox, plus a classifier), Allowlist, or Run Everything. The docs say Auto-review is not a security boundary. Plan Mode asks the agent to plan before editing, and agent review, diffs and PR creation happen in the Agents Window. For cloud agents, viewers are read-only by default, a person can take over the agent's remote desktop, and an admin can enable team follow-ups so teammates can continue a run. Projects delegate through a coordinator you chat with, and you open any agent it starts to talk to it directly.

## Protocols

| Protocol | Support | Evidence | Note |
| --- | --- | --- | --- |
| MCP | Yes (checked 2026-09-30) | [link](https://cursor.com/docs/mcp) | Client: the MCP page says Cursor connects to external MCP servers, and the cloud agent page says cloud agents use team-configured MCP servers over HTTP and stdio with OAuth support. |
| A2A | Unknown (checked 2026-09-30) | — | No mention of A2A or Agent2Agent found in the docs pages read (cloud agents, subagents, Projects, MCP, CLI ACP, SDK, pricing). Cursor is closed-source, so there is no repository to search. |
| AG-UI | Unknown (checked 2026-09-30) | — | No mention of AG-UI in the docs pages read, and Cursor is not listed in the AG-UI protocol README. Cursor CLI supports the Agent Client Protocol (ACP), which is a different protocol. |

## Best for

- Developers who want several agents on one repository at once, each isolated in its own worktree or cloud VM, managed from one window. ([shortlist](https://multiagentguide.top/best/parallel-coding-agents.md))
- Teams that start agents from Slack, GitHub, Linear or an API and want the results as pull requests with artifacts from the agent's own VM. ([shortlist](https://multiagentguide.top/best/coding-agents.md))
- Long-running work such as a migration or a multi-PR feature, using a Project coordinator that delegates to agents while you are away. ([shortlist](https://multiagentguide.top/best/coding-agents.md))
- Organisations that need admin controls, SSO, SCIM and audit logs around agent use (Teams and Enterprise plans). ([shortlist](https://multiagentguide.top/best/enterprise-governance.md))

## Not for

- Teams that need an open-source, self-hosted agent stack; Cursor is proprietary and its models are hosted (the SDK docs say inference goes through Cursor's hosted models in local mode too).
- Enterprise-plan users who want Projects: the Projects page says it is not available on Enterprise plans.
- Users who need a documented agent-to-agent protocol such as A2A; the docs describe MCP and ACP but not A2A.

## Quickstart

```sh
curl https://cursor.com/install -fsS | bash
```

Install not yet verified by this site.

```bash
# Cursor CLI (macOS, Linux, WSL); Windows PowerShell uses: irm 'https://cursor.com/install?win32=true' | iex
curl https://cursor.com/install -fsS | bash
agent --version

# Start an interactive agent session, or pass a first prompt
agent
agent "refactor the auth module to use JWT tokens"

# Non-interactive use: authenticate with an API key
export CURSOR_API_KEY="<your key>"

# Run the CLI as an Agent Client Protocol server over stdio
agent acp

# Desktop app: download it from cursor.com, then
# Command Palette -> Open Agents Window to run agents in parallel
```

### Common pitfalls

- The CLI install command is for macOS, Linux and WSL; native Windows uses the PowerShell command. Add `~/.local/bin` to PATH, then run `agent`.
- Cloud agents need an account admin to connect source control (GitHub, GitLab, Bitbucket Cloud or Azure DevOps) and read-write access to the repository and any submodules.
- Legacy Privacy Mode is not supported for Cloud Agents or Projects, because they store code in the cloud while running.
- The Agents Window worktree UI is only in the Agents Window; in the IDE use the Worktree Skills commands.
- The Cloud Agents API v1 is in public beta and may change.
- `@cursor/sdk` needs Node.js 22.13 or later; the Python package is `cursor-sdk`. Both need `CURSOR_API_KEY`.
- Do not symlink dependencies into a worktree; the docs recommend a fast package manager instead.

Official quickstart: https://cursor.com/docs/cli/installation

## Pros

- Parallel agents can be isolated locally in worktrees or remotely in cloud VMs, and moved between the two from the Agents Window. ([source](https://cursor.com/docs/agent/agents-window))
- Cloud agents can build, test and use a desktop and browser in their VM, and produce screenshots, videos and logs as artifacts with the pull request. ([source](https://cursor.com/docs/cloud-agent))
- Delegation is documented at several levels: built-in and custom subagents, cloud subagents, and Project coordinators. ([source](https://cursor.com/docs/subagents))
- TypeScript and Python SDKs and an HTTP API let other software start local or cloud agents. ([source](https://cursor.com/docs/sdk/typescript))
- Run Modes give three levels of approval for shell, MCP and Fetch calls, and Auto-review adds sandboxing and a classifier. ([source](https://cursor.com/docs/agent/security/run-modes))

## Cons

- Closed-source, subscription product; every plan includes a set amount of model usage and on-demand usage is billed after it is used up. ([source](https://cursor.com/pricing))
- Projects, the coordinator feature, is not available on Enterprise plans or with legacy Privacy Mode. ([source](https://cursor.com/docs/agent/projects))
- Cloud agents store code and environment data in Cursor's cloud while they run, and the Cloud Agents API v1 is a public beta that may change. ([source](https://cursor.com/docs/cloud-agent/api/endpoints))
- The docs state that Auto-review is not a security boundary and its classifier can allow calls you would have blocked. ([source](https://cursor.com/docs/agent/security/run-modes))
- SDK known limitations include that cloud agents reject local custom tools and that hooks are file-based only with no programmatic callbacks. ([source](https://cursor.com/docs/sdk/typescript))

## Alternatives

- [Conductor](https://multiagentguide.top/tools/conductor.md)
- [GitHub Copilot cloud agent](https://multiagentguide.top/tools/github-copilot-cloud-agent.md)
- [Devin](https://multiagentguide.top/tools/devin.md) ([Cursor vs Devin](https://multiagentguide.top/compare/cursor-vs-devin.md))
- [Claude Code](https://multiagentguide.top/tools/claude-code.md)
- [Codex CLI](https://multiagentguide.top/tools/codex.md)

## FAQ

### Does Cursor support MCP?

Yes, as a client. The MCP docs describe connecting Cursor to external servers, and cloud agents use MCP servers configured for the team over HTTP and stdio, with OAuth where needed.

### Can Cursor run agents in parallel?

Yes. The Agents Window is built for parallel agents, worktrees isolate each local task in its own checkout, and cloud agents run in separate VMs; the cloud agent page says you can run as many in parallel as you want.

### Is Cursor free?

There is a free Hobby plan with limited agent requests. Paid plans (Pro, Pro+, Ultra, Teams at $40 per user per month, Enterprise on request) add higher limits and cloud agents on Pro and above. The pricing page's structured data lists Pro at $20, Pro+ at $60 and Ultra at $200; checked 2026-09-30, and the page has a monthly and yearly toggle.

### How do subagents differ from Projects?

Subagents are child agents that the main agent launches inside one task, each with its own context window. A Project is a longer-lived coordinator agent that delegates whole tasks to agents on Cloud Agents and keeps shared context files.

## Sources

- [Cursor home](https://cursor.com)
- [Cursor docs home](https://cursor.com/docs)
- [Cursor pricing](https://cursor.com/pricing)
- [Cursor docs: Agents Window](https://cursor.com/docs/agent/agents-window)
- [Cursor docs: Worktrees](https://cursor.com/docs/configuration/worktrees)
- [Cursor docs: Cloud Agents](https://cursor.com/docs/cloud-agent)
- [Cursor docs: Subagents](https://cursor.com/docs/subagents)
- [Cursor docs: Projects](https://cursor.com/docs/agent/projects)
- [Cursor docs: TypeScript SDK](https://cursor.com/docs/sdk/typescript)
- [Cursor docs: Cloud Agents API](https://cursor.com/docs/cloud-agent/api/endpoints)
- [Cursor docs: CLI ACP](https://cursor.com/docs/cli/acp)
- [Cursor docs: Run Modes](https://cursor.com/docs/agent/security/run-modes)
- [Cursor docs: MCP](https://cursor.com/docs/mcp)
- [Cursor docs: CLI installation](https://cursor.com/docs/cli/installation)
- [Cursor docs index (llms.txt)](https://cursor.com/llms.txt)

## Unknown fields

A2A and AG-UI: searched the Cursor docs pages listed in note fields and the AG-UI README for both terms and found nothing. Cursor is closed-source, so no repository search was possible. The Python SDK (cursor-sdk) is documented at https://cursor.com/docs/sdk/python but not cited in sources. Pricing: Hobby, Pro $20, Pro+ $60, Ultra $200 and Teams $40 per user were read from the pricing page's JSON-LD and body text; the page shows a monthly/yearly toggle and the billing period of the JSON-LD figures was not stated, so treat non-Teams figures as list prices. Orchestration pattern 'supervisor' is based on the Projects coordinator page.

Corrections or removal requests: support@multiagentguide.top

---

Data as of 2026-09-30. Not affiliated with listed projects. HTML version: https://multiagentguide.top/tools/cursor