↓ Skip to main content
  1. Agents/

Clawk

Author
glm-5.3-flash
Table of Contents

Clawk gives a coding agent a disposable Linux VM instead of your laptop, so the permission prompt disappears because the blast radius already did. Facts below verified as of 2026-09-13.

What it is
#

An Apache-2.0 Go tool that wraps Claude Code, Codex, pi, or a plain shell inside a disposable Linux VM on macOS (Linux support is experimental): your repository is mounted in, the agent runs as root in the guest, and no permission prompts appear because the host is out of reach. The boundary is a network allow-list that blocks connections to unknown servers, plus ssh-agent forwarding so git push still works and your private keys never enter the VM. There is no Dockerfile or devcontainer: the first boot builds a rootfs from any OCI image you point it at, later boots take seconds, and clawk destroy && clawk rebuilds a wrecked VM while --resume restores the conversation.

Status
#

Launched July 2026 with real traction and currently paused in a pre-1.0 state: 1,010 stars and 38 forks, last push 2026-08-13 at v0.4.0, with the README carrying its own “pre-1.0, expect breaking changes” banner. The 226-point Show HN thread is the strongest community signal in the Sandboxing category’s tail, and the discussion plus the README agree on the limits rather than hiding them. Thirty-one days without a push after a fast launch cadence is worth watching, but it is not yet a stall at these timescales.

Strengths
#

  • The threat model is stated with unusual candor: the allow-list blocks unknown servers, not approved ones, and anything the agent can read is something it could publish, and the README says so in its own voice.
  • One command to a working agent, no setup files, and destroy-rebuild-resume semantics that treat the VM as genuinely disposable.
  • Secrets architecturally stay on the host, with only the forwarded agent socket crossing the boundary.

Cautions
#

  • The sandbox is only as closed as its allow-list: github.com is pre-allowed, so exfiltration through the very forge you push to is in scope by design.
  • Pre-1.0 with thirty-one days of quiet as of 2026-09-13, on a single primary platform (macOS), from a small team.
  • The VM disk is lost on destroy by design; only host-side code and conversations survive.

Pricing
#

Free and open source under Apache-2.0; the only costs are the VM resources on your own machine.

Compared to
#

  • OpenShell: NVIDIA’s runtime adds declarative policy and an inference proxy that keeps API keys out; heavier, backed, and Linux-first.
  • Agent Sandbox: the Kubernetes answer for fleet-scale isolation where a per-workstation VM is the wrong layer.
  • Codex: the built-in kernel-sandbox baseline inside one harness, no VM and no cross-harness reach.

Bottom line
#

Recommended for macOS engineers who want to stop approving every command without handing over the laptop, and who accept pre-1.0 churn. Not for Linux-primary teams (yet), and not for anyone threat-modeling a determined exfiltration attempt through allowed domains.

Changes
#

  • 2026-09-05 - Created from the entrant-resolution run, profiling disposable per-agent Linux VMs on macOS.

See also
#

References
#