You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 74b376a
Browse filesBrowse the repository at this point in the historyBrowse files
- Remove redundant Quick start section; /ai-stack:bootstrap now covers
starting services and registering MCPs
- Clarify that LINCE and OpenShell are two separate experimental sandbox
approaches: LINCE uses bwrap/nono + Zellij, OpenShell uses Podman +
network policy enforcement
- Note that LINCE is experimental and not yet integrated with OpenShell
- Note that OpenShell Vertex AI support is a temporary workaround
(NVIDIA/OpenShell#472)
- Document the tooling split: /ai-stack:sandbox bootstraps LINCE,
just sandbox-bootstrap bootstraps OpenShell
- Add reference to OpenKaiden (https://openkaiden.ai/) as a related
project for sandboxed AI agent execution
Co-Authored-By: Claude Sonnet 4.6 <[email protected]>
(gopls, pyright, typescript-lsp, rust-analyzer), Claude plugins (superpowers, context7,
36
+
atlassian, dev, track, quarterly), and registers MCP servers with Claude Code.
37
+
57
38
**Updating plugins:**
58
39
59
40
```bash
@@ -120,20 +101,37 @@ On first run, make any Google Workspace tool call — the server returns a click
120
101
URL. Complete the Google OAuth flow in your browser. Credentials are stored in
121
102
`~/.config/workspace-mcp/`and reused on subsequent runs.
122
103
123
-
## OpenShell sandboxes
104
+
## Agent sandboxes (experimental)
124
105
125
-
> **Note:** Running Claude Code in an OpenShell sandbox with Vertex AI currently requires
126
-
> a manual credential wrapper (`just sandbox`). This is a workaround until OpenShell adds
127
-
> native Vertex AI support (tracked in NVIDIA/OpenShell issue #472). Once that lands, the
128
-
> workflow simplifies to `openshell sandbox create -- claude` with a configured provider.
106
+
Running AI coding agents in isolated sandboxes is an active area. Two separate experimental
107
+
approaches are available here; neither is production-ready. See also
108
+
[OpenKaiden](https://openkaiden.ai/) — a desktop application that runs AI coding agents in
109
+
isolated sandboxes with enterprise governance controls.
129
110
130
-
`openshell/policy.yaml`grants sandbox network access to all MCP services in this stack.
131
-
Local services are reached via `host.containers.internal` (injected by the Podman driver).
132
-
Private IP ranges are explicitly allowlisted to override the built-in SSRF guard.
111
+
### LINCE (Zellij dashboard + bwrap/nono sandbox)
112
+
113
+
> **Experimental.** LINCE is a standalone toolkit and is not yet integrated with OpenShell.
114
+
> It uses `bubblewrap` (Linux) or `nono` (macOS) for filesystem/process isolation and
115
+
> Zellij for multi-agent session management.
116
+
117
+
Install or update via the plugin skill:
118
+
119
+
```
120
+
/ai-stack:sandbox
121
+
```
122
+
123
+
This installs `agent-sandbox`, the `lince-dashboard` Zellij plugin, and supporting scripts
124
+
from the local `lince/` checkout.
133
125
134
-
See `openshell/bootstrap.md` for full setup instructions and known issues.
0 commit comments