Cursor

Type: full-code · Vendor: Anysphere · Language: TypeScript (Electron IDE) · License: proprietary (closed-source) · Status: active · Status in practice: mature · First released: 2023

Links: homepage docs

A proprietary VS Code fork built around an integrated AI agent ('Agent') and tab-completion model ('Tab') that turns the IDE into the surface for tool-using coding agents.

Description. Cursor is the IDE; Agent is the embedded coding agent inside it. Agent runs autonomously with no fixed tool-call limit, edits files, runs terminal commands, searches the codebase via a semantic index, calls MCP tools, and creates checkpoints before significant changes. The IDE provides inline diffs, chat, apply-edit, and autocomplete (Tab). The codebase is indexed locally into encrypted vector embeddings that are kept in sync, and the index can be shared across team members. Cursor advertises multi-mode interaction: Ask, Agent, and Auto modes.

Agent loop shape. Agent runs an unbounded tool-use loop inside the IDE. Each tool call (read, search semantic-index, edit-file, run-terminal, MCP) is reasoned about and executed; results are observed and fed back. Checkpoints are taken automatically before significant changes so users can roll back. The Tab model is a separate, much-tighter loop: predict-as-you-type completions independent of Agent. Agent asks for approval before MCP tools by default; users can enable auto-run.

Primary use cases

IDE-embedded autonomous coding agent
semantic-vector codebase search
predictive autocomplete via Tab model
MCP-tool integration with approval gating

flowchart TD user[User prompt in chat] --> mode{Mode} mode -->|Ask| answer[Read-only Q&A] mode -->|Agent| agent[Agent loop] agent --> index[Semantic codebase index] index --> tool[Tool call: edit/terminal/MCP/web] tool --> approve{Needs approval?} approve -->|yes| user2[Approve in IDE] approve -->|no| exec[Execute] user2 --> exec exec --> ckpt[Auto-checkpoint] ckpt --> obs[Observe result] obs --> done{Task done?} done -->|no| agent done -->|yes| diff[Inline diff for review] diff --> apply[Apply edits] tab[Tab model] -.parallel.-> user

Key concepts

Agent (docs) — Autonomous IDE agent that edits, runs commands, and calls tools.
Tab (docs) — Predictive autocomplete model.
Codebase indexing → agentic-rag (docs) — Encrypted semantic vector index synced every 5 minutes.
MCP client → mcp (docs) — First-class MCP client supporting stdio/SSE/HTTP.
Checkpoints — Automatic codebase snapshots during Agent sessions.
Rules — .cursorrules and modern rules files steering Agent.

Cursor

Neighbourhood

Instantiates

Alternatives & relatives

Listed as alternative by (10)

References

Provenance