AI Agent Sandboxes

A structured, evidence-carrying dataset of sandboxes and adjacent approaches for running AI coding agents (Claude Code, Codex, etc.) with some form of isolation, network control, or credential mediation.

Started from the Show HN: Clawk thread and the official docs it cited; entries added since then are marked accordingly (see each entry's thread_status).

The data

Everything lives in ai-sandboxes.yaml:

• taxonomy — closed vocabularies for filterable facets (isolation boundary, network policy, credential mediation, threat tiers, etc.)
• sources — a registry of primary sources, referenced by id
• sandboxes — principal entries (sandbox runtimes, policy wrappers, remote workspaces, ...)
• adjacent — related tools and DIY patterns with a lighter schema

Every non-identity field is a claim, not a bare rating: it's either a plain scalar (shorthand for an official-evidence claim sourced from the entry's default_source) or a claim object with {value, ev, src, basis, derived_from}. ev: editorial claims must state their basis. See the conventions and evidence_levels keys in the YAML for the full rules.

Updating

Add entries following the existing shape, keep every claim's evidence traceable to a source in the sources registry, and bump meta.version with a short changelog note at the top of the file.

Roadmap

Plan is to stick a simple website on top of this soon — filterable by the taxonomy facets above, evidence links inline, "official claims only" toggle. For now it's just the YAML.

Shameless plug

One of the entries in here, byre, is mine. Its whole pitch is being comfortable — a local container that feels like your own machine, not a security appliance you have to fight. Give it a look!