Cursor、自社インフラで動くSelf-hosted Cloud Agentsを提供開始 Self-hosted Cloud Agents

Cursor has rolled out a self-hosted deployment option for its Cloud Agents, letting customers run the background coding agents on infrastructure they own and operate rather than on Cursor's managed environment. The change is aimed primarily at enterprises that need to keep source code, secrets, and execution telemetry inside their own security perimeter.

Cloud Agents are Cursor's asynchronous coding agents that take on tasks against a repository, run in parallel, and can carry out work such as drafting pull requests, performing larger refactors, or chasing down failing tests without the developer babysitting each step. Until now those agents executed inside Cursor-operated sandboxes, which simplified onboarding but made adoption harder for organizations with strict data residency, compliance, or network egress rules.

With the self-hosted model, the agent runtime can be deployed into a customer's own cloud account, typically a VPC or Kubernetes environment, so that repository contents and environment variables never leave controlled infrastructure. Because the network boundary is owned by the customer, agents can also more naturally reach internal services, private package registries, and bespoke build systems that would be awkward to expose to an external SaaS.

The move fits a broader pattern across the AI coding tools market. GitHub Copilot Enterprise, Sourcegraph Cody, Codeium, and agent-focused entrants like Cognition's Devin have all been layering on private deployment options, stricter data handling guarantees, and tighter identity integrations. For regulated sectors such as finance, healthcare, and the public sector, a deployment story where code does not transit a third-party cloud is often a hard procurement requirement rather than a nice-to-have. Cursor's self-hosted Cloud Agents read as a direct response to that pressure, and likely a prerequisite for some of its larger enterprise deals.

There are trade-offs worth flagging. Running agents in your own environment shifts part of the operational burden onto the customer: provisioning compute for parallel agent runs, managing autoscaling, monitoring long-running jobs, and keeping the agent runtime up to date with Cursor's releases. Model inference is another open question — depending on configuration, organizations may route to Cursor-hosted models, to their own keys with frontier providers, or potentially to private model endpoints, each with different latency, cost, and compliance profiles. It seems plausible that Cursor will continue to refine these options as more customers move from managed to self-hosted setups.

For teams already standardizing on Cursor, the announcement reduces a meaningful blocker around bringing agentic workflows into sensitive codebases. For the wider ecosystem, it is another datapoint suggesting that agent platforms, not just chat-style assistants, are becoming part of the enterprise developer tooling stack and are expected to meet the same deployment expectations as CI systems or code search before them.