Is it ready for production?
Section titled “Is it ready for production?”It runs one company’s production, on GKE, every day. It is also a 0.x project: the wire formats between components change, the API is not versioned yet, and the Firecracker it ships is a fork that is rebased on upstream daily. Plan on upgrading the whole deployment at once, control plane and host fleet together.
Do I need Firecracker?
Section titled “Do I need Firecracker?”For production, yes. Production isolation is Firecracker on Linux with KVM, and there is no container mode. Two other backends exist for development: Apple’s Virtualization framework on an Apple Silicon Mac, which gives you real microVM isolation on a laptop, and a subprocess backend with no isolation at all for iterating on the orchestration layer.
Which hosts can run the fleet?
Section titled “Which hosts can run the fleet?”Any Linux node with KVM. On managed Kubernetes that means a node pool with nested virtualization, which is Intel-only on every provider today: the C3 family on GKE, and either m8i or bare metal on EKS. Bare metal anywhere works too; the host agent does not care who owns the machine.
Can I use my own images?
Section titled “Can I use my own images?”Yes. A session image is any Linux image with /bin/sh, built with docker build and pushed to
any registry. There is no build tool to learn and nothing of engrams inside the image. Use a
glibc-based image: the bundled agent binaries do not run on Alpine.
Does anything happen without a person typing a prompt?
Section titled “Does anything happen without a person typing a prompt?”Yes. Automations start sessions from a schedule, a Slack mention, a pull request event, a Linear issue, or a signed webhook, and can prompt them, wait for them, run commands inside them, and post results to Slack or GitHub. Pull request review and Slack threads ship as automations you enable. There is no general approval block yet; a person is in the loop through the Slack thread, the dashboard, and the pull request review itself.
Which agents run inside?
Section titled “Which agents run inside?”Claude Code and Codex ship as harnesses. A harness is a bundle the host mounts into the VM at boot, so the agent runtime is upgraded fleet-wide without touching images.
Which models?
Section titled “Which models?”Whatever the harness supports with your key. The Claude Code harness talks to Anthropic with an API key or a personal Claude Code token; the Codex harness talks to OpenAI with an API key or a ChatGPT connection. Either can be pointed at OpenRouter instead, which fronts many providers behind one key.
How much does an idle session cost?
Section titled “How much does an idle session cost?”Nothing on the host. An idle session is a set of chunks in the blob store, deduplicated against the image’s base snapshot and every other session of that image, and a row in Postgres. A thousand sessions of a 4 GiB image take about 100 GiB of blob storage, not 4 TiB.
How fast is resume?
Section titled “How fast is resume?”Under 100 ms when the session resumes on the host that snapshotted it and the chunks are still in that host’s cache. One to two seconds on a different host, which is also the cost of migrating a session. A cold start from an image’s base snapshot is under a second and needs no pre-warmed pool. These are the numbers from the reference deployment; measure your own fleet before you promise them to anyone.
What is the license?
Section titled “What is the license?”engrams is free software under the GNU Affero General Public License v3.0.