Cloudflare's Workers AI Superintelligence team released updated Clairic and Clairic-flash, two extraction models that pull structured data from text, images, and PDFs, and made them open-weight. Both are hosted on Workers AI, and the full model weights are available on Hugging Face.
The new Clairic was redesigned to fill text and number fields in parallel, which the team says improved accuracy.
Clairic can now read PDFs as well as images, and handle up to four images or PDF pages in one request.
Responses are reported to come back in under one second, faster than typical LLM answers for similar tasks.
Cloudflare's self-reported internal evals are presented as directional only, with no public extraction benchmarks yet.
Cursor, the coding tool, now lets you hold real-time voice calls with AI agents (AI that splits up work and handles it on its own). You can ask for code changes out loud and hear how the work is progressing. It works in the agent panel, the development environment, the web, and the iOS app.
If the chat input is empty and the agent is idle, press the voice chat button to start a call.
On iOS, you can open a call by saying "Hey Siri, call Cursor." You can also mute or end the call from the lock screen.
The conversation continues while the agent is working, so you can ask it to check changes or report on progress.
You can choose from several voice options. With privacy mode turned on, the feature works without storing your data.
In Cursor, you can describe an app in words, and an AI agent will build it and deploy it. Finished apps can be shared with teammates through a single link. This feature is available on paid plans.
The agent builds the web app and also handles the database, login functionality, and hosting setup.
Before an app is made public, a security check looks for exposed sensitive information such as passwords.
The app goes live on the internet only after it passes the check and the user clicks publish.
Built apps can be shared with specific people or teams, and on personal plans they can also be opened so anyone can view them.
Cursor has officially released its Agent SDK. Developers can build agents as code folders, test them with evaluation tools, and deploy them to Cursor or their own servers. The tool is designed to make it easier for developers to build and run AI bots themselves.
Agents are built as a single project folder that includes instructions, tools, and technical procedures.
Bots can wake up and work based on events from GitHub, GitLab, Bitbucket, and Slack, or on a set schedule.
Evaluations, which check a bot's behavior with preset messages, can be run locally or in CI (automated build and test environments).
With a single command, developers can resend a real pull request as a fake webhook signal to test the bot.
Cursor's cloud agents can now create interactive canvases while they work. Finished canvases can be shared with a team through a link. A beta feature for defining and managing the cloud environments where agents run as code has also been released.
Cloud agents create canvases (screen layout documents) that users can interact with while the agent is working.
An Environment API (a tool for managing environments programmatically) is available in beta.
An environment can include up to 300 repositories, and large environments now load five times faster.
Origin has connected more than 20 developer tools to its platform. They cover deployment, CI, code review, security, monitoring, and project management. Partner apps include Cloudflare, Linear, Sentry, and Socket Security. Starting today, anyone who builds developer tools can submit an app to the marketplace.
You can choose apps for deployment (such as Vercel and Netlify), CI (build automation), code review, and security checks.
To build an app, you need to use the Origin API (the interface that lets outside programs exchange data with Origin).
An app starts out visible only to its creator. It then moves to public release and a review request.
Once an app passes review, it is listed in the marketplace. Submissions require a description tag and a demo video showing how the app is installed and removed.
Cursor's iPhone app now lets you view and send messages to AI agents (AI programs that handle tasks on their own) running on your own computer. The agents keep running on your computer, not in the cloud, and the app simply connects you to that computer.
In the app, tap your computer, then approve the connection request in the desktop app.
Your computer must be turned on and connected to the internet for this to work.
If you turn on the 'Keep this computer awake' setting, it won't fall asleep when you step away. A laptop must be charging and have its lid open.
Enterprise customers have this feature off by default, and an administrator must turn it on in the settings.
Cursor has released two bots that help with the final step of deploying code. Rollouts checks whether deployed changes work properly in each environment, and Security Review finds security flaws that could be exploited in every pull request (a proposed code change). Both bots are available right away on the Teams and Enterprise plans.
Rollouts posts a verification plan as a comment on each change, then checks logs and metrics after deployment and reports whether the change is normal, a regression, or still under review.
It checks each environment separately, so a change that looks fine in staging may still be flagged as a problem in production.
Security Review looks for injection attacks such as SQL injection, authentication bypasses, and exposed passwords, and reports the severity, the attack path, and a suggested fix.
Teams can also enforce their own coding rules by checking every pull request.
Cursor has released Projects, a feature for handing off large tasks such as building new functions or entire apps. Projects keeps context over work that spans months and divides the work among multiple AI agents (AI programs that split up tasks themselves). It is currently in beta and is open to all users starting today.
Projects runs on Cursor's own cloud computers, so work continues even when the user closes their laptop.
An orchestrator agent (an AI that plans work and assigns it to other agents) does not write code itself. Instead, it brings finished results to the user for review.
Each project has shared files, so testing methods and work approaches discovered by one agent can be used by the agents that follow.
When a Slack channel is connected, the orchestrator agent automatically assigns work whenever a bug report comes in.
Cursor has released a feature that lets AI agents run on a company's own machines. Code and passwords stay on internal devices and do not leave the company.
Teams can create a shared pool of machines. When a request comes in, an idle worker takes it on.
Machines with no requests are put to sleep and woken up again when follow-up requests arrive.
Agents can also run in cloud environments teams already use, such as AWS Lambda, Cloudflare, and Vercel.
On Linux and Mac, agents can view the screen and perform clicks, typing, and browser actions.
Cursor's cloud agents can now be used right away without connecting an external code repository such as GitHub. You can start building with a prompt, and if you like the result, save it to Cursor's own repository (Origin repo). You can preview the running app in your browser, and by connecting a Vercel account, you can also publish the output to a live address.
To begin, choose Start from scratch on the repository selection screen and make a request to the agent.
You can keep what you made in a repository by clicking the Create repo button, and choose its name and whether it is private or internal.
You can preview the environment where the agent runs directly in your browser.
A Vercel account is required to use the publishing feature.