All stories
engrams Ecosystem Community reports

engrams 0.11 lets coding-agent sessions move hosts without abandoning the run

The self-hosted Claude Code and Codex orchestrator adds durable session teleport and harness continuity, with important upgrade limits for existing installations.

By AITrending2 min read
Sources (3)

Why now The October 19 release of engrams 0.11.0 documents durable cross-host moves, swap-disk migration and explicit migration requirements for operators upgrading from 0.10.0.
engrams dashboard shows an agent conversation, a shared generated picture and a shell pane with the program used to produce it.
Maintainer-published session screenshot from the versioned engrams README. The example output and tool activity illustrate the existing dashboard; they are not an AITrending test or evidence of a successful cross-host migration. Source: Cortex Applications / engrams contributors

engrams wants moving an AI coding session between machines to be routine infrastructure work, rather than a reason to abandon the agent’s run. Its October 19 version 0.11.0 introduces durable session teleport, with a Postgres-backed state machine that the maintainers say survives a coordinator restart during the move.

The release also adds continuity for the agent harness: events are sequenced and acknowledged only after the coordinator stores them. A harness continuing from a snapshot can attach at its new host binding and finish its run. Live moves now include the guest’s swap disk, while host retirement waits for the coordinator to confirm that sessions have moved away. These are documented implementation claims; AITrending has not reproduced a live migration.

A self-hosted home for Claude Code and Codex

engrams runs agent sessions in separate Firecracker microVMs and exposes their transcripts through a dashboard, Slack or a CLI. Claude Code and Codex are bundled harnesses, with CLI versions 2.1.285 and 0.161.0 respectively in this release. Model access still comes from configured provider credentials; this is an orchestrator, not a new model or a provider subscription.

Production installation uses Terraform quickstarts and two Helm releases on GKE or EKS, with Linux/KVM hosts and cloud storage. Operators pay infrastructure and applicable model costs and configure secrets and allowed network destinations. The main code is AGPL v3; selected harness SDK crates are Apache 2.0. A local development fallback can run as plain subprocesses without isolation, so the production microVM claim does not apply to every development setup.

The upgrade has a one-way boundary

Existing operators should read the migration notes before updating. An idle session saved by 0.10.0 resumes through a disk-only cold boot: its files remain, but its running processes stop because the older snapshot lacks a swap manifest. That exception matters alongside the new continuity pitch.

The coordinator and all hosts must move together because their wire protocol changes. Host nodes must also be recreated for the new NBD device defaults. The release warns that database migration prevents rolling back to 0.10.0 and tells operators to back up both databases first. For teams willing to run this infrastructure, the attraction is explicit control over agent-session lifecycle; the cost is owning a still-changing 0.x system.

Related reading