Native Shell
Native Shell:AI 生成的 UI,在你的应用中原生呈现,覆盖每个平台。iOS、Android、macOS、Windows 与 Linux 上的应用、GPUI 应用与网页在运行时加载已签名、沙箱隔离的 TSX 模块并绘制 th
AI 翻译
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权,自行管理介绍,无需管理员批准。
登录并生成验证文件,在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件,即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录,将前往登录页面。登录后请回到此处,再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify,即可验证所有权并自行管理介绍。
一分钟完成验证Firecracker sandboxes on Kubernetes. Idle sandboxes sleep as snapshots in S3 and wake on any node. E2B-compatible API.
翻译准备中
取自仓库 README
正在加载完整 README。
查看完整 README ↗Sandboxes on Kubernetes, with an E2B-compatible API.
loam runs each sandbox as a Firecracker microVM, and keeps each one as a record in Postgres. A
sandbox runs on a long-lived worker pod (a nodelet, one per KVM node) only while it is in use.
When it goes idle, loam snapshots its memory and disk to S3-compatible storage and frees the
node, and the next request wakes it on whichever node has room, with its processes as they were.
You drive it through a native HTTP API, the loam CLI, Go and TypeScript clients, or the
official E2B SDKs, unmodified.
Project page and docs: https://loam.agentsh.it
E2B has a good API for agent sandboxes. Its own self-hosted setup runs on Nomad, Consul and Terraform. The Kubernetes projects that speak E2B's API run each sandbox as a pod, which ties every wake to the scheduler and keeps idle sandboxes holding memory. loam is for running Firecracker sandboxes on the Kubernetes cluster you already have, where a sleeping sandbox is a snapshot in object storage and costs nothing on any node.
Measured on the reference deployment, a three-node home k3s cluster on x86 bare metal with Garage as the snapshot store, using the base template (2 vCPUs, 2 GiB).
| What | Result |
|---|---|
| Resume a paused sandbox on the same node | 29 ms |
| Wake on another node, memory loaded lazily | 0.42 s |
| Fork from a memory checkpoint | 28 ms |
| Pause, memory and disk | 320 ms |
| Drain a node with two awake sandboxes | 13 s |
| Official E2B SDKs, Python and JS | 26 of 26 conformance steps |
The methodology is in the repo's test write-ups. docs/PLAN.md has the results
of each phase, docs/PHASE2-TEST.md and
docs/PHASE3-TEST.md are the multi-node and network boundary runs, and
docs/E2B.md describes the SDK conformance suite in e2e/e2b.
With a loam deployment that serves the E2B adapter, point the E2B SDK at it and use a loam project key as the E2B key.
pip install e2b
export E2B_DOMAIN=e2b.loam.example.com
export E2B_API_KEY=loam_...
from e2b import Sandbox
sbx = Sandbox.create()
print(sbx.commands.run("uname -a").stdout)
sbx.pause() # a snapshot in the store, the VM gone from its node
sbx.connect() # wakes it, on any node, with its processes
sbx.kill()
The native CLI does the same and more (exec sessions, port URLs, checkpoints, forks, egress rules, brokered secrets).
go install github.com/arugula-salad/loam/cmd/loam@latest
export LOAM_API_URL=https://api.loam.example.com LOAM_TOKEN=loam_...
loam create scratch
loam run scratch -- uname -a
loam pause scratch
loam run scratch -- uptime # wakes it first
To run loam yourself, start with docs/INSTALL.md. You need x86_64 nodes with KVM, Postgres, an S3-compatible bucket, an XFS data directory with reflinks on each node, and wildcard DNS and TLS for the sandboxes' URLs. To try it on one machine without a cluster, see "Locally" in docs/E2B.md.
loam is young. It was built in October 2026 and runs one deployment in production so far.
--memory./dev/userfaultfd.loam was written mostly by AI coding agents (Claude Code), working from Jake Gaylor's design and under his review. It replaced his earlier single-host Firecracker sandbox system and runs his own workloads in production, including browser games, hosted dev boxes, CI and an E2B endpoint.
See CONTRIBUTING.md. Report vulnerabilities privately as described in SECURITY.md.
Apache License 2.0, in LICENSE. The snapshot store is derived from e2b-dev/infra, also Apache-2.0. See NOTICE.