opendsec / README.md
zhyb99's picture
OpenDsec: self-hosted sandboxes implementing the DSec paper
07e1f12 verified
|
Raw History Blame Contribute Delete
3.63 kB
metadata
license: agpl-3.0
title: OpenDsec
emoji: 📦
colorFrom: blue
colorTo: gray
sdk: static
pinned: false
paper:
  - arxiv: 2609.22978
short_description: Self-hosted sandboxes for AI agents and RL

OpenDsec

OpenDsec is an open-source, self-hosted sandbox platform for reinforcement-learning rollouts and code-executing AI agents. It is an independent implementation of the architecture described in DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale (arXiv:2609.22978).

It is not DeepSeek code, not an official release, and not affiliated with or endorsed by the paper's authors. Where the paper leaves something open or the project chose differently, docs/SPEC.md marks each design point as either taken from the paper (with its section) or the project's own decision.

What it does

  • Backends: Firecracker microVM, container, and full QEMU VM behind one API.
  • Control plane: a stateless apiserver for auth, quota and placement; a Raft-replicated IAM; a watcher aggregating node health; an edge daemon per compute node that makes the final admission decision. Components use gRPC with mTLS, and the sandbox id encodes its edge, so no per-sandbox routing table is needed.
  • Lifecycle: pause to a memory snapshot, restore, and fork a running microVM with memory preserved. Same-node forks share pages; cross-node forks pull the snapshot from the source node.
  • Images: import from any OCI registry, compose layers without copying data, and pack a sandbox's changes into a new image.
  • Network: default-deny, with per-sandbox domain, CIDR and port rules enforced by nftables and a DNS proxy.
  • Compatibility: an E2B-compatible gateway, so existing E2B SDK clients can point at a self-hosted deployment.

Two deviations from the paper

  • A generic shared filesystem (NFS, CephFS or JuiceFS) stands in for 3FS.
  • Firecracker is the primary backend rather than one option among several.

Numbers

Every figure below is a measurement on the project's own hardware, not a figure from the paper. Conditions and method are in docs/SPEC.md.

Measurement Value
microVM cold start ~0.1 s
Restore from memory snapshot ~55 ms
Resume after pause ~50 ms
Burst of microVMs, one 384-thread node 5,120
Burst of containers, two nodes 8,192, no failures

Known limits

  • The image store's availability is whatever the chosen shared filesystem provides.
  • A host reboot loses that node's sandboxes, including paused microVMs.
  • QEMU PCI passthrough has not been tested on real devices.
  • The E2B gateway does not yet cover directory watching, signed file URLs, unauthenticated browser access to sandbox ports, or template building.

Links

Licensing

The server and tools are AGPL-3.0. The client parts (the Python SDK, the Go client, and the API definitions) are Apache-2.0, so programs using OpenDsec are not bound by the AGPL.

Citation

If you use OpenDsec in research, cite the paper it implements:

@article{dsec2026,
  title  = {DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale},
  year   = {2026},
  eprint = {2609.22978},
  archivePrefix = {arXiv}
}

"DeepSeek" and "DSec" are names of their respective owners. OpenDsec uses them only to identify the paper it implements.