Connectivity cloud (not cloud connectivity) company Cloudflare this month announced support for running Cursor Cloud Agents on Cloudflare Sandboxes. That move is designed to give software developers and platform engineering teams a way to run AI coding agents in secure, customer-controlled environments. As the realm of AI scaremongering shifts from ‘will it take our jobs’ to ‘okay it’s here, how do we use it safely?’, this news carries some key trends with it.
For the uninitiated, Cursor Cloud Agents are cloud-hosted AI coding assistants that independently plan, write code, run terminal commands and execute tests. They also handle long-running software engineering tasks in isolated remote environments. They are useful for deployments that have strict network requirements, because the code or services they house can’t be reached from outside an organisation’s network.
With Cursor Cloud Agents on Cloudflare Sandboxes, SpaceXAI users (i.e. coders who use Cursor, SpaceXAI’s automated code generation and support tool, acquired from Anysphere in June of this year) have another way to run AI agents in environments they control, while Cloudflare expands the infrastructure powering the next generation of agentic software.
What is a Cloudflare Sandbox?
By way of extra definition, Cloudflare Sandboxes allow developers to test and out run “untrusted code in isolated environments” so that AI agents, code interpreters, and developer tools have a secure place to execute, install packages, and interact with the filesystem.
The integration itself builds on Cloudflare’s work with other AI agent platforms, including Devin Outposts and Claude Managed Agents.
“Developers want to keep the tools they already use, while enterprises want control over where agent work runs and how it accesses code, systems, and secrets. Cloudflare Sandboxes are becoming a natural execution layer for that model: secure, isolated environments where agent work can run closer to a customer’s code, systems, and security requirements,” notes the company in a press statement.
Cursor Cloud Agents let developers assign coding tasks from the Cursor app, cursor.com, or the Cursor mobile app.
Cursor runs the agent loop
With Cursor Self-Hosted Machines, Cursor continues to run the agent loop, including inference, planning, and orchestration, while the agent gets work done on a customer-selected worker.
In SpaceXAI’s model, a worker is the customer-operated machine or environment that executes agent tasks; it is separate from Cloudflare Workers, Cloudflare’s serverless developer platform. With Cloudflare Sandboxes, that self-hosted worker can run in a sandbox environment in the customer’s Cloudflare account.
According to Dane Knecht, chief technology officer at Cloudflare, developers “keep the Cursor workflow they already use” and they still start and manage agents from Cursor. Additionally, Cursor still routes the work and streams results back to the user.
What changes is where the work happens.
Tool calls, including terminal, filesystem, and browser actions, run inside customer-controlled Cloudflare sandbox environments, which matters for teams with strict requirements around where code, build caches, and secrets live. For organisations using Cursor Self-Hosted Machines across different execution environments, Cloudflare Sandboxes provide another controlled option for running agent workloads.
“Developers want powerful AI tools that fit naturally into their workflows, and enterprises need those tools to run in environments they control,” said Dane Knecht, Chief Technology Officer at Cloudflare. “Cloudflare Sandboxes give teams the freedom to use the AI tools they prefer while giving organisations a secure, isolated place to run agent work. Bringing Cursor Cloud Agents to Cloudflare Sandboxes is another step toward making Cloudflare the execution layer for the next generation of agentic applications.”
Cursor Cloud Agents run via self-hosted machines and use an outbound connectivity model. A Cursor worker runs the Cursor CLI and opens a long-lived outbound HTTPS connection to Cursor’s backend, where agent tool calls are sent over that connection. Cursor does not need to open an inbound connection into the customer’s network; teams can use the quickstart guide to configure a Cursor worker and connect it to their environment.
SpaceXAI’s goals (apart from colonising Mars)
“SpaceXAI’s goal is to make agents useful wherever developers work,” said Toni Adams, Sr Director of Partnerships at SpaceXAI. “Self-Hosted Machines let teams keep Cursor workflows across desktop, web, and mobile while routing work to infrastructure they operate. Cloudflare Sandboxes give those teams an enterprise-grade option for running agent workloads in a controlled environment.”
The integration supports Cursor’s Self-Hosted Machines workflow for individual developers and teams. Developers can connect a single worker through My Machines, while enterprise teams can use Cursor self-hosted worker pools as named routing targets that allow new agent chats to wait until an available worker claims them.
Teams can create pools for different execution environments, then use pool orchestration to watch demand, start worker capacity when needed, and release it when sessions end.
During self-hosted operation, repositories, build caches, and secrets stay on the customer’s machines. File chunks read by the model during inference, along with Cloud Agent artefacts such as screenshots, videos, and log references, are uploaded so they can appear in pull requests and dashboards. Teams that want to integrate self-hosted machine status or pool routing into their own systems can also use the Cloud Agents API.
What should developers think next?
In terms of what software application development professionals should think here, this integration represents a shift toward sovereign AI development i.e. what we means is… rather than sacrificing enterprise security policies to use autonomous coding assistants, engineers can now run AI agents within their own security boundaries. Programmers will no doubt like the fact that proprietary code, build caches, and sensitive API secrets never leave their organisation’s network.
The promotional show t-shirt here (if there was one) would probably say something like: “Developers can now scale agentic workflows securely without altering how they write code.”