Cursor

What Cursor is and who develops it

Cursor is a development environment and coding-agent platform from Anysphere, which operates under the Cursor brand. Its editor originated as a Visual Studio Code fork, while the product has expanded toward complete tasks: understanding a repository, preparing a change, executing it and presenting evidence for review. It has been part of SpaceX since 14 August 2026.

Distinguish the editor, agent and model. The editor provides navigation, files and development tools; the agent organizes actions; the model interprets context and proposes those actions. Cursor can use proprietary models and models from other providers depending on access. The original editor’s license does not make all Cursor services or models open source.

Its main audience builds or maintains software. For an OSH team, its role concerns digital resources supporting prevention: portals, data transformations, forms and libraries. It is not a dedicated prevention register and does not decide which preventive measure should be applied.

Reference documentation: Accel — Cursor founders · Contrary — Founding story · Cursor — SpaceX acquisition completed · Cursor — Downloads.

Origins, financing and ownership changes

Anysphere was founded in 2022 by four MIT classmates: Michael Truell, Sualeh Asif, Aman Sanger and Arvid Lunnemark. The company established itself in San Francisco. In 2023 it raised an eight-million-dollar seed round led by OpenAI Startup Fund. The project focused on integrating AI throughout the programming environment rather than only providing isolated suggestions.

• 2024: a $60 million Series A was announced in August.
• 2025: a $105 million Series B was announced in January, followed by a $900 million Series C in June at a $9.9 billion valuation.
• November 2025: the $2.3 billion Series D valued the company at $29.3 billion. It reported exceeding $1 billion in annualized revenue, following $100 million in January and $500 million in June.
• 2026: the partnership announced with SpaceX in April preceded the completed August acquisition.

Annualized revenue describes a business run rate reported at a particular date, not profit or audited revenue for an entire financial year. Inclusion in the April 2025 Forbes AI 50 provides context about prominence rather than an independent assessment of code accuracy.

Lunnemark left his operating role in October 2025 to work at Integrous Research. Accounts of recruitment in 2025 described initial exercises without AI and two-day on-site projects. These are reported historical practices, not a universal current policy for every role.

Reference documentation: Accel — Cursor founders · Contrary — Founding story · Cursor — Series A · Cursor — Series B · Cursor — Series C · Cursor — Series D · Cursor — SpaceX acquisition completed · Forbes — AI 50 announcement · Svenska Dagbladet — Lunnemark departure · Bridge Chronicle — Historical hiring account.

How Cursor works with a software project

Cursor is a development environment with AI assistance. Its Agent mode can explore a codebase, edit multiple files, run commands and prepare tests. Useful input combines a specific request with project context; output includes file changes requiring review, rather than just an explanation in chat.

Edits are applied during the work and can be inspected in a diff view. Cursor also documents checkpoints for restoring files; restoration does not delete conversation messages. Maintain version control and distinguish local changes, completed checks and published code: these are different development states.

Reference documentation: Cursor Docs — Agent mode.

Editor evolution and acquisitions

Version 1.0, dated 4 June 2025, introduced Bugbot, expanded background agents and added Jupyter and MCP developments. Cursor 2.0 added a parallel-agent interface and the Composer model in October. The Cursor 3 Agents Window, introduced in April 2026, provided a workspace for supervising multiple jobs while retaining the traditional editor.

Acquisitions reflect different parts of development. Supermaven joined in November 2024 for its autocomplete technology. Its standalone wind-down was announced in November 2025, retaining autocomplete inference for existing users while ending agent conversations.

Koala and Resourcely teams joined in 2025 with enterprise-adoption and security expertise. This did not turn Cursor into Koala’s CRM. The Graphite acquisition agreement was announced in December, with its code-review product continuing independently. Firetiger joined in August 2026 with a focus on production software reliability. Acquiring a team does not imply that every feature of its earlier product is included in every Cursor plan.

Reference documentation: Supermaven — Joining Cursor · Supermaven — Sunsetting · Koala — Acquisition announcement · Cursor — Graphite agreement · Cursor — Firetiger · Cursor — 1.0 changelog · Cursor — 2.0 changelog · Cursor — Agents Window.

Local work, cloud execution, rules and version control

The desktop is available for Windows, macOS and Linux. There are also CLI and web agent interfaces; mobile access includes an iOS/iPad application and an installable Android PWA. These are ways of directing work, not guarantees that code executes physically on the device used to write the request.

Cloud agents run in virtual machines with their repository and environment. Local tasks depend on files, commands and credentials available on the computer. Git worktrees separate parallel changes, but two solutions can still conflict during integration even if neither overwrote the other’s files.

Rules define project, user or team conventions. Project rules use MDC files in .cursor/rules; AGENTS.md offers another instruction mechanism. A rule to preserve each site’s identifier guides the agent, while actual separation between sites requires application controls and tests.

A verifiable workflow starts from a known branch, reproduces the problem, makes a bounded change, checks the result and presents the diff. Record process decisions separately: what counts as pending, which date takes precedence and who may access records. The agent needs these definitions to avoid resolving business ambiguity by guessing.

Reference documentation: Cursor — Agents Window · Cursor — Cloud Agents · Cursor — Rules · Cursor — Downloads.

Models, Composer and evaluation context

Cursor separates its product from the models it runs. Composer 2, introduced in March 2026, used Kimi K2.5 as its base, adding continued pretraining and reinforcement learning on programming tasks. Incomplete initial attribution prompted criticism and a company acknowledgment.

Composer 2.5 documentation describes a model specialized for extended programming tasks, with standard and fast variants. A faster variant changes interaction and cost; it does not establish that fewer reviews will always be needed. Cursor’s published CursorBench results concern its own evaluation, not certification of an OSH portal.

Compare models using the same project version, request and acceptance cases. Count failed attempts, reviews and discarded changes too. Changing model, context and available tools simultaneously makes it difficult to establish which factor produced an improvement.

Reference documentation: Cursor — Composer 2 technical report · Cursor — Composer 2.5 · Economic Times — Composer attribution.

Bugbot, Origin and review before production

Bugbot reviews pull-request diffs and reports potential problems. One operational detail matters: its check is neutral by default when findings exist. Requiring that check does not necessarily block changes with unresolved findings. A review skipped because of size must not be counted as a favorable review either.

Origin provides Git hosting and mirroring in early beta. When hosting a repository it can become the source of truth; when mirroring GitHub, GitHub retains that role. Before moving a project, document where branches are created, changes reviewed and access administered.

Rollouts and Security Reviewer were introduced for Teams and Enterprise in September 2026. The former connects changes with deployment signals; the latter looks for software-security issues. Monitoring errors after an OSH portal update supports availability but does not verify its preventive recommendations.

Define what review authorizes: commenting, proposing a fix and merging code are different actions. An automated proposal can provide a second reading without replacing the maintainer’s acceptance criteria.

Reference documentation: Cursor — Bugbot · Cursor — Origin · Cursor — Rollouts and Security Reviewer.

Plans, consumption and continuing projects

Current documentation integrates Cursor into The X Subscription and lists Free, Plus, Premium, Super, Ultra, Teams and Enterprise. It distinguishes a weekly first-party model pool from separate prepaid third-party consumption on eligible plans. Legacy plan conditions also remain documented. A tutorial referring to “Pro” may describe an older subscription.

Evaluate three costs: seats, execution and review. One long task can consume more than many brief questions; a shared account also makes work and permission attribution harder. Check the actual account’s access to models, cloud agents, Bugbot and repositories rather than inferring it from the plan name.

Projects, introduced in beta in September 2026, organizes extended work through a coordinator, agents and shared context. It can sustain an effort spanning multiple tasks. For a preventive-content library, the objective might be batches of link and translation fixes, each retaining affected pages, verification and an acceptance owner. Continuity does not remove the need for a clear completion criterion.

Reference documentation: Cursor — Models and pricing · Cursor — Projects.

Documented incidents and adoption lessons

In April 2025, the Sam support bot invented a single-device policy; the company apologized. The episode distinguishes generated information from a confirmed contractual condition.

In July 2025, Cursor acknowledged unclear communication of June changes to its Pro plan and offered refunds for unexpected charges during the affected period. Moving from request counts to metered consumption requires explaining what the subscription includes and what can incur extra costs.

On 29 October 2026, Cursor reported altered Anthropic.claude-code extension versions 2.1.312 and 2.1.313 in its marketplace. The malicious code was not developed by Anthropic. Cursor removed the packages and rebuilt the marketplace; its notice states that uninstalling may not remove persistence.

These events concern different problems: support answers, pricing communication and extension distribution. For an organization, the practical decision is to review generated work alongside installed components and environment permissions. A provider’s correction of one incident does not remove the need to manage future updates.

Reference documentation: Cursor — Pricing explanation · Ars Technica — Sam support incident · Cursor — Extension security notice.

Applications related to occupational safety and health

The following uses are Sabentis editorial proposals for teams developing resources that support prevention:

• Improving forms: review labels, error messages and keyboard navigation to reduce completion mistakes.
• Preparing reports: write transformations of authorised data that preserve identifiers and detect incomplete records.
• Maintaining digital content: locate untranslated text or broken links in a library of preventive materials.

Cursor helps build or maintain software. It does not independently provide the methodology for determining risk or turn a generated form into a validated assessment. Functional specifications should come from process owners.

Hypothetical example: correcting an outstanding-actions filter

An internal portal displays actions from several sites. Its outstanding filter appears to include closed actions when a date is missing. The team wants to correct the logic without changing real data.

1. Explain the available statuses and what “outstanding” means. Supply fictional data containing an open action, a closed one, a cancelled one and an incomplete record.
2. Ask for the filtering function to be located and its behaviour explained before changing it.
3. Define the expected result for each record. Incomplete data should be identified rather than assumed closed.
4. Review the diff and run checks on the filter and screen. Confirm that totals and labels agree.
5. Also test a user restricted to one site. A visual correction should not alter access controls for other sites.

The verifiable result is a bounded change with reproducible input and output cases. A reply saying “fixed”, or a screen that appears correct, does not replace those checks.

Expand the test set to ten actions: four open, three closed, two cancelled and one without a reliable status. The pending view should show four, with a separate data-quality issue; neither seven nor five is the expected result. Changing a closed action’s date must not accidentally reopen it. Also retain a test rejecting another site’s records at the server, rather than merely hiding rows on screen.

Privacy, context and agent limitations

Cursor’s documentation distinguishes Privacy Mode from processing: with this mode, code is not used for training, but AI features send prompts and code context to providers. It should therefore not be interpreted as fully local execution.

Restrict context to what is necessary and use fictional data to reproduce errors. Do not include credentials or personal case files in examples. Review commands, dependencies and affected files; a change can compile while still calculating an indicator incorrectly. Restoring files should not be confused with reversing operations performed in external services.

Reference documentation: Cursor Docs — Agent mode · Cursor Docs — Privacy and data.

When to choose Cursor and how to compare it

It is an option for teams wanting to work on existing code in an editing and review environment. It requires the ability to interpret changes, maintain the project and check subsequent behaviour.

Codex provides another approach to assisted project work; v0 is relevant when exploring a web interface through instructions. Compare them on a representative task and measure correctness, clarity of changes and review effort alongside generation speed.

Frequently asked questions about Cursor in OSH

Can it create an application ready for preventive use?

It can help develop one, but suitability requires specifications, testing, content review and maintenance. A working interface does not establish valid preventive logic.

Does Privacy Mode mean nothing leaves the computer?

No. According to the documentation, it protects against training use, but AI features process context through providers.

What should the person accepting a change review?

The scope of modified files, test cases, permissions, data used and behaviour expected to remain unchanged.

Reference documentation: Cursor Docs — Privacy and data.