Continue.dev¶
Continue is an open-source VS Code coding agent. In
agent mode it runs shell commands and edits files on your behalf. By default
those commands run on your host. Point Continue at IronClaw's sandbox_exec
tool and the agent runs untrusted, AI-generated commands inside an isolated
gVisor sandbox instead.
Why sandbox Continue?¶
A coding agent executes code the model wrote, not code you reviewed. One bad command can read your credentials, exfiltrate a token, or delete files. IronClaw contains each command in an ephemeral gVisor container: no host filesystem, no unexpected egress, destroyed after the session.
- See it stop real escape attempts: nivardsec.com/containment
- Read the isolation benchmark: Containment benchmark
1. Install ironctl¶
Verify:
2. Edit the config file¶
Continue uses YAML config. Open (or create) ~/.continue/config.yaml and add
the ironclaw entry under mcpServers:
If you already have other MCP servers, add the ironclaw entry alongside them
in the same mcpServers list.
MCP runs in agent mode
Continue only exposes MCP tools in agent mode. Switch the Continue chat
to Agent before expecting sandbox_exec to appear.
3. Reload¶
Save the file. Continue reloads its config automatically. If the tool does not appear, run Developer: Reload Window from the VS Code command palette.
What sandbox_exec does¶
- Runs commands in an ephemeral container managed by IronClaw
- Uses gVisor for kernel-level isolation -- the AI cannot read your host files or make unexpected network calls
- Container is destroyed after the session ends
Troubleshooting¶
ironctl: command not found -- Make sure ironctl is on your $PATH.
Run which ironctl to check. If you installed via Homebrew, run
brew link ironsecco/ironclaw/ironclaw.
Tool does not appear -- Confirm Continue is in Agent mode, check the config for YAML indentation errors, and reload the VS Code window.