Devin can now run on your own hardware

Cognition launched Devin Outposts this week, a deployment option that splits its Devin coding agent into two parts: the reasoning and planning loop, which continues to run in Cognition’s cloud, and command execution, file edits, and repository access, which now happen on infrastructure the customer controls. Supported targets include a Mac mini, a GPU server, a virtual machine inside a private network, or a Kubernetes cluster sitting next to internal services. Cognition built deployment templates with Cloudflare, Daytona, E2B, Modal, Namespace, and NVIDIA Brev to cover common environments.

Cognition frames this as solving a specific enterprise adoption blocker: coding agents need to touch a company’s actual codebase and internal services, but running that execution entirely inside a third-party cloud is a nonstarter for many security and compliance teams. Outposts keeps the expensive model inference in Cognition’s infrastructure while letting the sensitive, environment-specific work, the part that touches real credentials and real systems, stay on hardware the customer already controls and audits.

This ships the same week Cognition acquired the maker of the personal-messaging agent Poke, and both moves point the same direction: Cognition is building out infrastructure options around a single “always-on agent” core rather than shipping one narrow product. For teams evaluating coding agents, Outposts is worth a look specifically if data residency or network isolation has been the blocker, since it changes where code execution happens without changing which agent you’re using. See what running a coding agent remotely actually requires before committing to any self-hosted deployment.