Sandstorm is a self-hosted Rust platform that fits thousands of agent sandboxes onto machines by pausing idle ones, snapshotting them to disk, and restoring them quickly when traffic arrives.
Sandstorm’s bet is that agent sandboxes idle most of the time, so a tiered pause-snapshot-restore lifecycle, not bigger machines, is what makes a sandbox fleet affordable.
What it is #
One binary, sandstormd, serves as the master, the node agent, and the web console, with the master holding desired state in SQLite and placing sandboxes across nodes.
Sandboxes run in one of two backends, gVisor containers or Firecracker microVMs, each sandbox in its own network namespace, and both backends can pool in one fleet.
The lifecycle has four tiers: hot (running), frozen (paused, cold pages in zswap, about 0.1 s back to hot), snapshot (checkpointed to disk, about 0.2 s back), and detached (volume only, cold start).
Nodes join by address and pairing code or dial out from behind NAT, a per-node gateway wakes a sandbox on incoming traffic, and nodes refuse new work under memory pressure so the host stays reachable.
Apache-2.0, by DeepLethe (whose own site, deeplethe.com, is a Chinese-language company page), created 2026-05-11 under the name forkd and since renamed (the old path redirects; a forkd-action repository remains).
Status #
Active and newly released: 3,270 stars, 293 forks as of 2026-10-31 (GitHub API), 11 open issues and PRs as of 2026-11-18, pushed 2026-11-18, with v0.6.0 (2026-11-01) the newest release after the five-month v0.5.3 stall.
v0.6.0 shipped hours after the release-preparation pass that fixed the packaging (the release binary had been built against glibc 2.38 and would not start on Ubuntu 22.04 or Debian 12, and the Docker image had quietly stopped building): gVisor becomes the default backend, so a node needs no KVM while Firecracker stays opt-in per node behind --backend firecracker, master, node, and web console collapse into the one binary, the control plane runs over mutual TLS in both directions with each sandbox’s own token checked at the gateway, host groups add fleet placement, the design is recorded as ADRs, the console was rewritten in TypeScript and React, and the official E2B SDKs work against Sandstorm.
The README now publishes three rounds of fleet measurements (2026-10-28 through 10-31) with methodology and raw result links, plus a self-run comparison chart against CubeSandbox, microsandbox, Kata, Docker, and bubblewrap.
The post-v0.6.0 fix train continued through 2026-11-18 (gateway-activity retries after failed node reports, output isolation between concurrent runsc calls, snapshot-tag encoding in the Python SDK, admission enforcement for pinned placement, and revoked-node rejection after long-poll wakes).
Zero independent footprint: a Hacker News search returns no story about this project as of 2026-11-18, and every benchmark on the README, including the runtime comparison chart, is self-published and self-measured. The name collides with Sandstorm.io, the long-archived self-hosted app platform, which still owns search results for the word.
Strengths #
- The lifecycle is the feature: published wake latencies per tier (about 0.1 s frozen, about 0.2 s from snapshot) aim the design at idle memory, the actual cost driver of sandbox fleets.
- Real isolation boundaries, not wrappers around them: gVisor and Firecracker are first-class backends with their own optimization passes, documented with before-and-after measurements in the repo’s OPTIMIZATIONS.md.
- Fleet plumbing in one binary: SQLite desired state, NAT-dialing nodes, wake-on-traffic gateway, and self-protecting nodes.
Cautions #
- The five-month release stall ended on 2026-11-01, but v0.6.0 is a breaking cut (the old forkd-controller removed, per-sandbox tokens required, a Node 20 SDK floor), so pre-0.6 adopters face a migration, and the 0.x cadence is one release in five months.
- Every performance number is self-measured, now published with methodology and raw results but with no third-party benchmark or write-up found anywhere.
- Single-vendor project with a Chinese-language company presence and no English community channel surfaced, so roadmap visibility and support are thin.
- The Sandstorm name fight against Sandstorm.io for search and mindshare is a permanent tax.
Pricing #
Does not apply: the platform is open-source and self-hosted, and no hosted tier is on offer.
Compared to #
- Microsandbox: a local-first embeddable microVM runtime for sandboxes on one machine; choose Sandstorm when the fleet spans machines and idle cost, not sandbox startup, is the problem.
- CubeSandbox: Tencent’s self-hosted, E2B-compatible microVM sandbox service; it is API-first where Sandstorm is lifecycle-first with a web console and placement.
- E2B: the hosted Firecracker API that defines the category baseline; choose it for zero ops, and Sandstorm to own the machines underneath.
Bottom line #
Recommended for engineers running self-hosted agent-sandbox fleets on their own machines, who can accept a days-old 0.x major and self-measured numbers. Not for teams wanting a hosted API, a long release record, or an independently reviewed product.
Changes #
- 2026-10-31 - Created.
- 2026-11-01 - Volatile refresh: recorded the v0.6.0 release-preparation pass on main (2026-11-01), which repairs the release packaging (a binary built against glibc 2.38 that would not start on Ubuntu 22.04 or Debian 12, and a Docker image that had stopped building), bumps core and SDKs to 0.6.0, and cuts no tag yet, with the repository pushed 2026-11-01.
- 2026-11-08 - Reworded the CubeSandbox See-also line to drop a banned term (now “closest in kind”); no factual change.
- 2026-11-13 - Rotation re-verification: every cited URL re-fetched and alive (repository, README, OPTIMIZATIONS.md, the releases page, the maker’s site, the HN search still returning zero stories), and the release record corrected: v0.6.0 shipped 2026-11-01, hours after the release-preparation pass the note had recorded as tagless, making gVisor the default backend with Firecracker opt-in per node, collapsing master, node, and console into one binary, adding the mutual-TLS control plane with per-sandbox tokens, host groups, ADRs, a TypeScript/React console, three published rounds of fleet measurements, and E2B SDK compatibility, with open issues and PRs at 14 as of 2026-11-13 and the Status, Cautions, and Bottom line updated off the unreleased reading.
- 2026-11-18 - Rotation re-verification: every cited URL re-fetched and alive (repository, README, OPTIMIZATIONS.md, the releases page, the maker’s site, the HN search still returning zero stories); the post-v0.6.0 fix train recorded through 2026-11-18 (gateway-activity retries, runsc output isolation, snapshot-tag encoding, admission enforcement, revoked-node rejection), with open issues and PRs moving from 14 to 11 and the repository pushed 2026-11-18.
See also #
- Microsandbox - the local-first microVM runtime this note’s lifecycle economics answer at fleet scale
- CubeSandbox - the self-hosted microVM sandbox service closest in kind
- E2B - the hosted baseline every self-hosted entry here is measured against
- Overalls - the other self-hosted Firecracker platform in the category
References #
https://github.com/deeplethe/sandstorm - repository, feature set, backends, and lifecycle claims
https://raw.githubusercontent.com/deeplethe/sandstorm/HEAD/README.md - tier table, latency chart, key features, and naming
https://raw.githubusercontent.com/deeplethe/sandstorm/HEAD/docs/OPTIMIZATIONS.md - the per-backend optimization evidence
https://github.com/deeplethe/sandstorm/releases - the stalled release train (v0.5.3, 2026-07-22)
https://deeplethe.com - the maker and its Chinese-language company presence
https://hn.algolia.com/api/v1/search?query=deeplethe%20sandstorm&tags=story - the zero-story Hacker News footprint as of 2026-10-31