We posted an arXiv report on training Composer 2—from continued pretraining on Kimi K2.5 through large-scale RL in realistic Cursor sessions—and on CursorBench, public benchmarks, and infrastructure.
作者: Sasha Rush語言: en-US
索引內容摘要
Blog / researchMar 27, 2026·researchA technical report on Composer 2Sasha Rush · 3 min readTable of Contents↑Continued pretraining and RLReal-world evaluation with CursorBenchPerformanceInfrastructureWe posted to the arXiv a technical report on the training of Composer 2, our coding model for agentic software engineering. The report covers the full training process, from continued pretraining on an open base model, Kimi K2.5, through large-scale reinforcement learning, with a focus on closely emulating the real Cursor environment. #Continued pretraining and RL Composer 2 is trained in two phases: continued pretraining on a data mix that emphasizes code to deepen the base model's coding knowledge, followed by large-scale reinforcement learning to improve end-to-end agent performance. We find that reducing pretraining loss improves downstream RL performance, with better base knowledge reliably translating into a better agent. Composer 2 RL training occurs in realistic Cursor sessions with the same tools and harness the deployed model uses, applied to a problem distribution that reflects the full range of what developers ask Composer to do. We find that RL training improves both avera…
首次發現: 最近檢查: 內容更新:
文章站內連結
Aug 14, 2026·CompanyCursor is now a part of SpaceXCursor Team · 2 min read
Blog / companyAug 14, 2026·companyCursor is now a part of SpaceX2 min readCursor has officially been acquired by SpaceX. This completes the acquisition process that started in April, when we announced our partnership with SpaceXAI to accelerate our model training efforts. In the years since we started Cursor, better models have steadily expanded what people can build. Cursor has gone from completing the next few lines of code to building AI teammates that you can give real work to. Together with SpaceX, we will push that ambition further. We will have access to the largest fleet of GPUs in the world, giving us the compute to build stronger models that are also more economical to run. This means we can provide customers with more capable models at lower cost. Grok 4.6, which we released Wednesday, provides an early look at what we can now build together. SpaceX is building the computing capacity needed to scale intelligence far beyond what exists today. Cursor will be one place where that intelligence becomes useful. For us, that opens a much larger horizon than the one we started with, while keeping the work familiar. We still want to help people with ambitious ideas spend less tim…
首次發現: 最近檢查: 內容更新:
文章站內連結
Aug 12, 2026·ResearchIntroducing Grok 4.6Cursor Team · 3 min read
Built for long-running agents and more ambitious interactive and visual work.
作者: Cursor Team語言: en-US
索引內容摘要
Blog / researchAug 12, 2026·researchIntroducing Grok 4.63 min readTable of Contents↑Training Grok 4.6Turning ambitious ideas into working projectsSafety and capabilitiesGet started with Grok 4.6Today we are releasing Grok 4.6 together with SpaceXAI. Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work. It stays with complex tasks across many steps, whether researching a topic, analyzing information, working across a codebase, or turning an idea into a polished application or work artifact. Grok 4.6 achieves frontier intelligence across several agentic coding and knowledge work benchmarks. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, which is a composite score of nine benchmarks. Grok 4.6 is available today in Cursor and Grok Build. We’re offering 2x included usage inside Cursor and Grok Build for the first week. #Training Grok 4.6 Grok 4.6 underwent a longer supplemental training run than Grok 4.5, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe. This produced a stronger foundation…
AI code review that catches real bugs. Bugbot detects the hardest logic bugs with a low false positive rate.
語言: en-US
索引內容摘要
{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://cursor.com/bugbot","name":"Bugbot","description":"AI code review that catches real bugs. Bugbot detects the hardest logic bugs with a low false positive rate.","url":"https://cursor.com/bugbot","applicationCategory":"DeveloperApplication","operatingSystem":"macOS, Windows, Linux","publisher":{"@type":"Organization","@id":"https://cursor.com/#org"}}{"@context":"https://schema.org","@type":"FAQPage","mainEntity":[{"@type":"Question","name":"How does Bugbot integrate with my existing workflow?","acceptedAnswer":{"@type":"Answer","text":"Bugbot finds bugs directly in GitHub. It automatically reviews PRs, comments on potential issues, and provides fixes directly in your Cursor editor or through our Background Agent."}},{"@type":"Question","name":"Can I customize how Bugbot does reviews?","acceptedAnswer":{"@type":"Answer","text":"Yes! Bugbot Rules allows you to define custom coding standards, enforce proven approaches, and set up project-specific guidelines to match your team's needs."}},{"@type":"Question","name":"How accurate is Bugbot's bug detection?","acceptedAnswer":{"@type":"Answer","text":"Bugbot uses …
Delegate repetitive work to always-on agents. Set up Cursor Automations that run on schedules or event-based triggers to build, maintain, and fix software.
語言: en-US
索引內容摘要
Save time by eliminating repetitive tasks.Set up agents that run automatically based on schedules or event-based triggers.Build your first automation→Trusted by engineering teams atCreate your self-driving codebaseAgents that autonomously build and maintain software.Improve code qualityAgents autonomously investigate incidents, detect and fix bugs, and generate tests.Speed up time to mergeAgents independently analyze PR risk, triage and fix CVEs, and respond to reviews with fixes.Improve codebase knowledgeAgents regularly share data updates, produce codebase digests, and update docs.See value immediatelyStart building with popular templates.Daily digest of repo changesSet up an agent that posts a daily digest summarizing notable repository changes and new risks introduced.Build from template ↗Automatic test coverageCreate an automation to review every new commit and add tests for high-risk logic to improve test coverage.Build from template ↗Bugs in Slack → PRs with fixesSet up an agent to monitor Slack for bug reports, conduct investigations, and return PRs for simple fixes.Build from template ↗Assess PR risk → Assign reviewersCreate an automation to review code changes, determine …
Built to help you ship, right from your terminal. Same commands, any environment.
語言: en-US
索引內容摘要
{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://cursor.com/cli","name":"Cursor CLI","description":"Built to help you ship, right from your terminal. Run Cursor agents in the terminal, GitHub Actions, and automations with the same commands in any environment.","url":"https://cursor.com/cli","applicationCategory":"DeveloperApplication","operatingSystem":"macOS, Windows, Linux","publisher":{"@type":"Organization","@id":"https://cursor.com/#org"}}CLIShip code with agentsRight from your terminal.Same commands, any environment.Plug into your setup anywhere.curl https://cursor.com/install -fsS | bashInteractive demo showing agent. The interface is displayed over a scenic painted landscape wallpaper, giving the demo an artistic backdrop.agentGet CLI> agentCursor Agent~/anysphere/research · main→Plan, search, build anythingGPT-5.6 Sol Extra High Fast/ commands · @ files · ! shellAlways access the latest modelsFrontier models from Anthropic, OpenAI, Gemini, Cursor, and more, at your fingertips.Cursor Agent~/cursor/cursor-web→→/model Auto/model Grok 4.6/model Opus 5/model GPT-5.6 Sol High Fast/model Fable 5.1 Max/model Gemini 3.1 Pro/model Composer 2.5Use in your…
{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://cursor.com/pricing","name":"Cursor","description":"Choose the Cursor plan that's right for you. Free for individuals, with paid plans for power users, teams, and enterprises.","url":"https://cursor.com/pricing","applicationCategory":"DeveloperApplication","operatingSystem":"macOS, Windows, Linux","publisher":{"@type":"Organization","@id":"https://cursor.com/#org"},"offers":[{"@type":"Offer","name":"Hobby","price":"0","priceCurrency":"USD"},{"@type":"Offer","name":"Pro","price":"20","priceCurrency":"USD"},{"@type":"Offer","name":"Pro+","price":"60","priceCurrency":"USD"},{"@type":"Offer","name":"Ultra","price":"200","priceCurrency":"USD"},{"@type":"Offer","name":"Teams","price":"40","priceCurrency":"USD"}]}{"@context":"https://schema.org","@type":"FAQPage","mainEntity":[{"@type":"Question","name":"What is the right plan for me?","acceptedAnswer":{"@type":"Answer","text":"We recommend Pro+ for daily agent users, and Ultra for agent power users. The Teams plan is recommended for professionals collaborating with others, and larger organizations that need invoicing, pooled usage, or advanced security should ch…
Run Cursor from your browser or phone. Cloud agents control their own computers to build ambitious software.
語言: en-US
索引內容摘要
The third era of AI software developmentA third era of AI software development is emerging as autonomous cloud agents take on larger tasks over longer timescales.Ideas · Feb 26, 2026
Use Cursor everywhere you workA unified workspace for agents across every surface. @font-face { font-family: 'CursorIcons16'; src: url('/marketing-static/fonts/cursor-icons-16.woff2') format('woff2'); font-weight: normal; font-style: normal; font-display: block; } @keyframes fadeIn { from { opacity: 0 } to { opacity: 1 } } @keyframes fadeSlideUp { from { opacity: 0; transform: translateY(2px) } to { opacity: 1; transform: translateY(0) } } @keyframes fadeSlideRight { from { opacity: 0; transform: translateX(-2px) } to { opacity: 1; transform: translateX(0) } } @keyframes barGrow { from { transform: scaleY(0) } to { transform: scaleY(1) } } @keyframes tilePopIn { from { opacity: 0; transform: scale(0.7) translateY(3px) } to { opacity: 1; transform: scale(1) translateY(0) } } Acme Research Dashboardbuild a dashboard to visualize our findingsExplored 12 files, 4 searchesOn it. I'll build the dashboard using your theme config, wire up the research data, and add interactive charts with public access controls.Worked for 14m 22sProcessed screen recordingDone! Here's a walkthrough of the dashboard.Acme LabsSummaryBuilt the interactive dashboard with realtime charts, data from Snowflake, an…
Join Cursor Cafes, meetups, hackathons, and workshops happening in dozens of cities around the world. Connect on Discord, Reddit, or the Forum.
語言: en-US
索引內容摘要
Join the global Cursor community.Explore Cursor meetups, cafes, hackathons, and workshops around the world.Meet the communityParticipateTweet by @victorianoi: I spent the day today at the Cursor Cafe in Madrid. I really liked the format. They simply rent a spacious coffee shop where you can bring your laptop and meet lots of other people who use Cursor to program with AI (they give you $50 in AI credits) I ran into many friends (Madrid@victorianoiTweet by @rasaljaya: 3 days to go Cursor Meetup Colombo 🇱🇰 This was Cafe Cursor in January… same energy, but way bigger this time⚡ Nick Miller is all set to be here 👀 See you all soon. #ColomboTech #SriLankaTech #AIBuilders #lkaColombo@rasaljayaTweet by @benln: Cafe Cursor in MedellínMedellín@benlnTweet by @benln: Cafe Cursor in MilanMilan@benlnTweet by @benln: Cafe Cursor in PraguePrague@benlnTweet by @mwanje_ronnie1: 😎 how it endedKampala@mwanje_ronnie1Tweet by @AHadzibabic: Lovely vibes at Cafe Cursor in Novi Sad! Thanks to everyone who attended and contributed to an amazing event. More to come soon. 🥂Novi Sad@AHadzibabicTweet by @GeorgeOlivine: Meeting some amazing builders and founders here at #cafenbo BlackieNairobi@GeorgeOlivin…
Cursor is transforming the way we build software. Millions of developers at the world's most ambitious companies already code with Cursor, but we have much more work to do.We obsess over talent to an unusual degree and are designing a company that's a haven for self-motivated individual contributors. This work demands our best.If this sounds exciting, we'd love to hear from you.Open rolesAll teams↓All teamsSalesCustomer SuccessUser OpsMarketingEngineeringRevenue OperationsDesignSolutionsPeoplePartnershipsAll locations↓All locationsAPJAustinAustraliaBengaluruEMEAGermanyIndiaJapanLondonMumbaiNew DelhiNew YorkPhilippinesRemoteSan FranciscoSingaporeWashington, D.C.Account Executive - CommercialSales·Full-time·SingaporeApply →Account Executive, Commercial Sales·Full-time·LondonApply →Account Executive, Commercial - San FranciscoSales·Full-time·San FranciscoApply →Account Executive, Commercial Expansion - New YorkSales·Full-time·New YorkApply →Account Executive, Commercial Expansion - San FranciscoSales·Full-time·San FranciscoApply →Account Executive, Geo Enterprise - ANZSales·Full-time·AustraliaApply →Account Executive, Geo Enterprise - New YorkSales·Full-time·New YorkApply →Account Exe…
Build from anywhere via the Cursor mobile app on iOS. Launch agents when ideas strike, get notified when work is ready for review, and merge PRs on the go.
語言: en-US
索引內容摘要
{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://cursor.com/mobile","name":"Cursor Mobile","description":"Work with Cursor on the go from the native mobile app on iOS.","url":"https://cursor.com/mobile","applicationCategory":"DeveloperApplication","operatingSystem":"macOS, Windows, Linux","publisher":{"@type":"Organization","@id":"https://cursor.com/#org"}}MobileBuild from anywhereWork with Cursor on the go from the native mobile app on iOS.iOSiOS @font-face { font-family: 'CursorIcons16'; src: url('/marketing-static/fonts/cursor-icons-16.woff2') format('woff2'); font-weight: normal; font-style: normal; font-display: block; } @keyframes fadeIn { from { opacity: 0 } to { opacity: 1 } } @keyframes fadeSlideUp { from { opacity: 0; transform: translateY(2px) } to { opacity: 1; transform: translateY(0) } } @keyframes fadeSlideRight { from { opacity: 0; transform: translateX(-2px) } to { opacity: 1; transform: translateX(0) } } @keyframes barGrow { from { transform: scaleY(0) } to { transform: scaleY(1) } } @keyframes tilePopIn { from { opacity: 0; transform: scale(0.7) translateY(3px) } to { opacity: 1; transform: scale(1) translateY(0) } } Follow up...The b…
Sep 10, 2026 · ChangelogCursor Projects Today we're launching Projects in Cursor. Projects lets you take on larger bodies of work, such as a feature, a migration, or a full app. It maintains context over months of work, delegates tasks to thousands of subagents, and performs recurring work without being prompted. Access Projects from the left-hand nav. The coordinator agent in a project doesn't write code itself; it plans the work, delegates it to agents that implement it, and brings the finished work back to you to check. Coordinators create and manage agents on your behalf, running as many in parallel as the work needs. #Powered by Cloud Agents A Project runs on its own computer in the cloud, so closing your laptop doesn't stop it. When something needs testing on your machine, the coordinator spins up a local agent to run it there. #Shared context You shouldn't have to onboard an agent every time you start a task. Each Project maintains a set of files that sync across every cloud and local machine its agents use. Agents add research and artifacts, along with what they learn about the codebase and how you prefer work to be done. If one agent figures out how to test a service, for …
AI code review only helps with real codebase context. What changed as agents grew, what to measure, and how to run Bugbot and Agent Review.
作者: Cursor Team語言: en-US
索引內容摘要
Guides / code reviewAug 12, 2026·code reviewAI code review: more context, fewer bugs8 min readTable of Contents↑What changed about code reviewQuestions before launching AI code reviewWhat is AI code review?How is it different from linting or CI?Does AI code review replace human review?What makes AI review noisy?What should we measure?When should review run: locally, on the PR, or both?Do we need the review tool to live in the same product as the agent?How to run AI code review in CursorAutomate the routing and sign-offKeep review close to the codeCode review is often slow and uneven: diffs wait in queues, and feedback depends on who is online. AI code review only helps when the reviewer has real codebase context. AI code review works best when the reviewer already has the full repo, recent changes, tests, and a clear set of rules. AI code review with this context catches problems a standalone bot staring at a single diff will never see. Review is most useful as part of the same system that produced the change. #What changed about code review Coding agents did not invent long pull requests, but they made the old review math fail faster. Developers working with coding agents are ship…
Model routing sends each request to the model that fits the job, so routine work stops billing at frontier prices. How it works, what to measure, and how to run it.
作者: Cursor Team語言: en-US
索引內容摘要
Guides / modelsAug 12, 2026·modelsModel routing: right model, right job, right price7 min readTable of Contents↑Why model routing matters nowThe questions worth answering before you routeWhat is a model router?How does routing decide which model to use?Should routing be automatic or should engineers still choose?Does routing hurt quality?What should you measure?How does routing handle caching and switching models?How to run model routing in CursorMatch the model to the jobMost teams pay frontier prices for work that does not need a frontier model. Someone picks the strongest model they can find, sets it as the default, and every request after that runs at the top rate. The spend climbs faster than the quality of what ships. Model routing is the fix. Instead of sending every request to one model, a router looks at each request and picks the model that fits. Simple work goes to fast, cheap models. Hard, long-horizon work goes to the frontier models. Model routing is turning into a standard layer in how teams work with LLMs. The gap between the cheapest capable model and the most expensive one is huge, and most of what you do all day sits at the easy end. Route it well and you keep th…
Engineering teams are past their AI budgets and still accelerating. Seven questions on cheaper models, cloud agents, metrics, budgets, and coaching.
作者: Cursor語言: en-US
索引內容摘要
Guides / StrategyAug 12, 2026·StrategyMaximizing return on intelligence8 min readTable of Contents↑1. Should I use cheaper models?2. Should I be using open weight models?3. Should I use cloud agents? Aren't they expensive?4. What metrics should I be tracking?5. Should I set spend budgets per engineer?6. How should I be reviewing my AI spend?7. How should I be coaching my team?By David Pan Toward the end of last year, agentic engineering finally started working. And once it worked, everything grew at once. More engineers picked up the tools. They started using them for more tasks. And each task got more expensive as a result of more thinking and tool calls. Those three things multiply, which is how so many engineering teams watched their AI spend go on a J curve. These teams will tell you the spend has been worth it. They are shipping faster and getting real value. But worth it or not, they're flying past their 2026 budgets, and nearly every customer we talk to now has some version of the same mandate: keep accelerating AND get disciplined about the cost. Easier said than done. Here's my take on some of the questions that come up most often. #1. Should I use cheaper models? Everyone…
MCP connects coding agents to your databases, tickets, docs, and error trackers. Which servers to start with, and how to add one in minutes.
作者: Cursor Team語言: en-US
索引內容摘要
Guides / integrationsAug 12, 2026·integrationsMCP for coding agents: how agents plug into your stack8 min readTable of Contents↑Why agents need real tool accessQuestions before you connectWhat is an MCP server?How is MCP different from a plugin or a normal API integration?Does the agent run these tools on its own?What makes a good MCP server versus a noisy one?Best MCP servers for Cursor, by use caseHow to add an MCP server in CursorBuild your own MCP serverWire it into the stack you already useAn agent can already work in your repo: search code, edit files, run commands, and browse the web. But it can't see your error tracker, database, tickets, or docs. Until those are connected, it stalls on anything that lives outside the editor. The Model Context Protocol (MCP) is an open standard that connects an agent to outside tools and data. Connect an MCP server and the agent queries the database instead of asking you for the schema. It opens the ticket instead of describing what the ticket should say. That access is what separates an agent your team leans on from one writing code in isolation, and it makes the coding agent MCP strategy one of the first things teams need to solve for. Th…
Prompting is like web design. Let’s call it prompt design, and build better tools for it.
作者: Arvid Lunnemark語言: en-US
索引內容摘要
Blog / ideasJun 11, 2023·ideasPrompt designPrompting is like web design. Let’s call it prompt design, and build better tools for it.Arvid Lunnemark · 6 min readTable of Contents↑Priompt v0.1: a first attempt at a prompt design libraryCaveatsObligatory last noteI’m usually averse to the common habit of trying to find old-world analogues to new-world phenomena. So please bear with me as I commit that exact sin: let me present the case for why prompting should be called prompt design and be likened to web design. I view prompting as communicating with a time-constrained human. While LLM-specific techniques are certainly helpful (most notably chain-of-thought), I’ve found that one of the best ways to improve performance is to just have extremely clear and high-quality instructions, similarly to how clarity and conciseness helps real humans understand better too. Prompting-as-clear-communication makes prompting sound like writing. Most of the prompting I am doing, however, is parametric: I have a number of input variables, and need to dynamically adapt my prompt to those. Hence, prompting-as-clear-communication-with-dynamic-input feels like the most accurate characterization. What other…
A new model and inference method for high-accuracy full-file edits at 1000 tokens/s.
作者: Aman Sanger語言: en-US
索引內容摘要
Blog / researchMay 14, 2024·researchEditing Files at 1000 Tokens per SecondAman Sanger · 9 min readTable of Contents↑Evaluating Prompted RewritesSpeed MeasurementsDiff ModelsTrainingSynthetic DataModel TrainingSpeculative editsFuture DirectionsFrontier models such as GPT-4o struggle on large edits, with problems of laziness, inaccuracy, and high-latency. This is a weakness visible in coding agents. Accurately editing hundreds of lines can take multiple model calls, at times trapping the agent in an infinite loop. Even small, isolated edits are plagued with bugs. Worst of all, existing models are slow at large edits, breaking the programmer out of flow. We've trained a specialized model on an important version of the full-file code edit task called fast apply. Difficult code edits can be broken down into two stages: planning, and applying. In Cursor, the planning phase takes the form of a chat interface with a powerful frontier model. Applying the change to the current file should be straightforward and instant. A toy example of a change we want to 'apply'. It cannot easily be copy/pasted since it sketches out the change over a full class Our fast-apply model surpasses GPT-4 and GPT…