What the installer sets up
Pre-write checks fire on every Write, Edit, and MultiEdit. Layer 1 runs locally. Layer 2 runs against your FlowRail workspace.
Pre-bash checks fire on dependency installs (npm install, pip install) so a flagged package is blocked before it installs.
Skills are registered with Claude Code so it can open reviews, request verification, check dependencies and render lineage on demand; scanner-finding triage arrives in v1.
// .claude/settings.json
{
"hooks": {
"PreToolUse": [
{ "matcher": "Write|Edit|MultiEdit",
"hooks": [{ "type": "command",
"command": "FLOWRAIL_MCP_URL=https://api.flowrail.ai npx --yes -p @flowrail/hook flowrail-hook pre-write" }] },
{ "matcher": "Bash",
"hooks": [{ "type": "command",
"command": "FLOWRAIL_MCP_URL=https://api.flowrail.ai npx --yes -p @flowrail/hook flowrail-hook pre-bash" }] }
]
}
}Day-to-day
Open Claude Code as you always have. When a check fails, the agent receives a structured block with the predicted threat and a suggested fix; it patches its own write and retries. The cycle takes seconds. You get involved only when neither layer can resolve the situation.
Other agents
Cursor, Aider, Copilot Workspace, and Devin are on the roadmap. The protocol layer (MCP + skills) is agent-agnostic; what's in flight is parity for each agent's extension format. Using one of these and want early access? Reach out.