danger-full-access
--sandbox danger-full-access + escape hatch
How to use it
danger-full-access is exposed by the Agent Rack MCP server. Add the server to your MCP client (Claude Desktop, Cursor, Windsurf and others), and the danger-full-access tool becomes available to the model automatically. See the full listing for setup details and every tool this server provides.
Install Agent Rack
npx agent-rack setupOther tools in Agent Rack (31)
No parameters. Lists every configured agent and whether its binary is on $PATH.
Same execution parameters as agentrun (agent, prompt required; workspace, mode,
An agentreview call returns
Delegating any task to a sub-agent through agent-rack
claude mcp add agent-rack -- node <resolved-bin-path> start. --scope maps directly to Claude Code's own -s local\
codex mcp add agent-rack -- node <resolved-bin-path> start (global only — codex has no project-scope flag).
Merges an mcpServers.agent-rack entry into Cursor's mcp.json, plus copies agent-rack's two guidance skills into Cursor's skills/ directory. --scope user (default) writes to ~/.cursor/; --scope project writes to <project>/.cursor/ instead.
Merges an mcpServers.agent-rack entry into Claude Desktop's config (macOS only).
Agents
runs once the process exits, reducing all accumulated events into a
builds the argv for a single run, given the prompt and an optional mode.
/lineend may be 0 for whole-file, deleted-file, or architectural findings.
opencode
Merges an mcp.agent-rack entry into opencode's config ($OPENCODECONFIGDIR, else $XDGCONFIGHOME/opencode, else ~/.config/opencode) — note this target uses a different config shape ({ type: "local", command: [...] }) than the others. Global only.
agent-rack install --target opencode
identical to agentrun but with agent pre-filled.
is called on every stdout/stderr chunk as it streams in; return zero or more
--sandbox read-only (enforced)
claude, codex