The free beta is open.
Automated security review of every file Claude Code writes, checked before it reaches disk. Free while we're in beta, with a one-time $5 LLM credit to get you started.
What FlowRail does
FlowRail is a security guardrail for AI coding agents. It hooks into Claude Code and checks every file the agent writes or edits before it reaches disk: a local regex net for leaked secrets (sk_live_, AKIA, ghp_ and friends) plus an LLM verifier running nine guardrails on the code itself.
The point is timing. Scanners and CI catch problems after the code exists — in a diff, in a pipeline, in a report someone has to read. FlowRail sits at generation time, so the agent hears about the problem inside its own loop and can fix it before you ever see the file.
Who it's for
Indie developers using Claude Code to build real things who aren't security people. If you're shipping a side project or an early product and you couldn't say offhand what a CSV formula injection or an SSRF sink looks like, that's exactly the gap this covers. You shouldn't have to become a security engineer to let an agent write your code.
What you get in the beta
Automated security review of every file Claude Code writes: the pre-write verifier checks each write and edit against nine guardrails before it lands.
Dependency checking: npm and pip installs are gated against four live npm-registry signals, your design review's approved dependency set, and OSV.dev known-CVE checks for npm and PyPI.
Secret detection: hardcoded keys and tokens are caught locally before they're written, and every catch lands in an auditable lineage you can review later.
It's free. There are no paid tiers yet. Every beta account gets a one-time $5 LLM credit — the verifier calls a model on every check, and that credit covers it while you kick the tires.
How to install
Run npx @flowrail/init in your project. It wires up the Claude Code hooks and MCP server; you sign up, mint an API key, and export it as FLOWRAIL_API_KEY. That's the whole setup. Claude Code is the only agent we wire and test today.
What's not here yet
No paid tiers, no BYOK, no enterprise features. Verification is a single-model semantic check today, and the default fail mode is fail-open: if FlowRail can't complete a check, you get a visible warning and the write proceeds (opt into fail-closed with FLOWRAIL_FAIL_MODE=closed).
One thing to know before you install: your code is sent to api.flowrail.ai and one LLM provider for the check. It's processed in-memory and never persisted — we store metadata, SHA-256 fingerprints, and verdicts, wiped within 24 hours — but it does leave your machine. A no-egress mode is on the roadmap, not in the beta.
Try it, tell us what breaks
FlowRail is built by a security researcher with confirmed critical vulnerabilities across the GitHub, GitLab, Elastic, Cosmos, and Mozilla bug bounty programs. The guardrails encode the classes of bugs that actually get paid out, not a generic checklist.
It's a beta, and we want it beaten on. Install it, point Claude Code at something real, and tell us what it caught, what it missed, and where it got in your way.