Engrams is a self-hosted orchestrator for AI coding agents that runs each agent inside its own Firecracker microVM, snapshots the VM when it goes idle, and restores it on demand. Sessions are built as plain OCI images plus a small harness; once enabled, the system boots images on hosts in a user’s cloud, streams transcripts and tool calls to a dashboard, Slack or CLI, and enforces strong isolation with KVM and host-side outbound filtering. Key selling points are full control of data and keys (everything stays in your account), aggressive deduplication of disk and memory via hash-keyed chunks, near-instant restore times (sub-100 ms on the same host, a second or two across hosts), and built-in product features - dashboard, CLI, connectors, automations, and task pipelines - so it behaves like an integrated sandbox product you operate yourself.
The architecture splits product surface and scheduling from execution: an orchestrator handles users, integrations and automation, a coordinator schedules sessions and evicts idle VMs, and many host agents dial in to run VMs. Storage uses content-addressed blobs with lazy paging (NBD + userfaultfd) so snapshots only write changed chunks and base images are stored once across sessions. Deployment targets Kubernetes with Terraform and Helm for production; a single-machine dev flow supports macOS virtualization or Firecracker on Linux. It supports Claude Code and Codex harnesses, isn’t a hosted service or model provider, and is a 0.x project running production at one company.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.