Features
Seven checks in. One verdict out.
Sentrail is not a checklist and not a wrapper around npm audit. It collects timestamped evidence from your database, repositories and deployments, and will only call a project ready when all of that evidence exists and is current.
Required checks
What does Sentrail read during a security check?
Sentrail reads and timestamps evidence from your database, repositories, and deployments into append-only scan receipts. Original provenance is preserved while permitted finding lifecycle fields can change, and the launch verdict is recomputed from current qualifying evidence.
Connectivity
Confirms the connectors an audit depends on are actually active, so no verdict is ever computed with a blind spot.
connectivity
Database RLS
Reads every non-system table and its policy count to catch tables with Row Level Security off, and tables with RLS on but zero policies.
rls_scan
Supabase advisors
Pulls Supabase's own security and performance advisor lints — exposed views, definer functions, unsafe function search paths, auth configuration.
supabase_advisors
Code scan
Searches your connected repositories for service-role keys and other credentials sitting in client-reachable paths or public-prefixed env vars.
code_static_scan
High-risk file review
Locates and reads the security-critical files — auth handling, middleware, webhooks, env plumbing, migrations — and reviews what they actually do.
repo_high_risk_read
Dependency audit
Resolves your package manifests and checks each dependency against the OSV vulnerability database for known advisories.
dependency_audit
Deployment posture
Classifies deployment environment variables by risk and inspects deployment protection settings on the hosting side.
deployment_check
Verdict logic
How are the four security states decided?
Sentrail decides security states by validating if evidence is current within a 24-hour freshness window. Critical or high severity findings block launches outright, while lesser findings cap the score to 'needs review'.
Every required check has fresh evidence and there are zero open critical or high findings.
Evidence is complete and nothing critical or high is open, but medium and low findings are waiting on a human decision.
Evidence is complete and at least one open critical or high finding stands between you and launch.
One or more required checks has never run, failed, or has gone stale — so no score is reported at all.
Evidence is only counted as current inside a freshness window — 24 hours by default. Outside it, the check is stale and the verdict drops to insufficient evidence. Findings that are still open at critical or high severity block a launch outright; medium and low findings hold the verdict at needs review and are capped so they cannot quietly sink the score.
"Ready" certifies required connectors are active, every required check has evidence from the last 24h (default), and there are zero open critical/high findings across code, database, Supabase advisors, dependencies, and deployment posture. It does not include the deep Semgrep/Gitleaks/OSV scanner, DAST-lite, the LLM-wiring audit, or a human review — those are separate, optional checks.
Beyond the seven
Deeper checks, kept out of the verdict on purpose.
These run on demand. They are useful, they are real, and they are excluded from the ready boundary so that boundary stays precise.
Deep code scan
Runs Gitleaks, OSV-Scanner and Semgrep against a commit in an isolated scanner service, with secrets redacted before triage.
DAST-lite
Probes a running deployment for a small set of externally observable issues, such as missing security headers.
LLM wiring audit
Reviews how the codebase wires up language models — key handling, prompt construction and tool exposure.
Remediation
A fix is a proposal until a human says otherwise.
Sentrail writes fixes. It does not apply them behind your back, and it never merges anything.
- 01
Proposal
The agent writes a fix proposal — SQL, or exact file contents — against the finding. Nothing is written to your database or repository at this stage.
- 02
Awaiting approval
The proposal parks in an approval queue with its diff, its blast radius, and the finding it closes. Sensitive actions cannot skip this queue.
- 03
Human approval
A reviewer on your side approves or rejects. Approval is recorded with who approved it and when.
- 04
Draft PR only
On approval Sentrail opens a draft pull request on its own branch. It re-checks the base commit first and aborts if the branch moved underneath it.
- 05
You merge
Sentrail never merges. Every change lands through your normal review and CI, and the re-verification scan runs after the merge you performed.
Draft pull requests only. Merges stay with you and your CI.
Agent behaviour
An operator that has to show its work.
Runs the checks itself
Ask for an audit and the agent triggers the checks that are missing or stale, rather than reasoning over old evidence.
Refuses to guess
If required evidence cannot be collected, the answer is insufficient evidence — not a confident verdict.
Specialist subagents
Database, code, secrets, posture and recon work is delegated to focused agents with only the tools that job needs.
Sensitive actions queue
Anything that would change your database or repository is written to an approval queue instead of being executed.
Continuous scanning
The same rigour on every commit.
- Webhook
A push or pull request on a monitored repository queues a scan run.
- Scanners
An isolated scanner service runs Gitleaks, OSV-Scanner and Semgrep against the commit tarball.
- Redaction
Matched secrets are fingerprinted and redacted before any finding reaches a language model.
- Triage
Findings are ranked and labelled — true positive, likely false positive, needs review — with reasoning attached.
- Report
Results post back as a pull request comment and a check run. Blocking merges on criticals is opt-in per repository.
Connectors
What Sentrail connects to today.
Supabase
via OAuthSchema, RLS policies, advisor lints and auth configuration — read through the Management API.
GitHub
via GitHub AppRepository reads and evidence-backed review of connected source, with remediation guidance.
Vercel
via OAuthEnvironment variable risk classification and deployment protection posture.
29 tools for your coding agent.
Mint a scoped token and point Claude Code or Cursor at the Sentrail MCP server. Read-scoped tokens can inspect findings; scan-scoped tokens can also trigger audits. Fix proposals still land in the approval queue.
- endpoint
- https://sentrail.dev/api/mcp
- transport
- Streamable HTTP
- auth
- Bearer token, scoped read or scan
Control mapping
Findings mapped to the frameworks you get asked about.
OWASP ASVS
4.0.3
Application security verification requirements
SOC 2
TSC 2017
Trust services criteria
ISO/IEC 27001
2022
Annex A controls
PCI DSS
v4.0
Requirement references
Sentrail maps findings to control references. It does not assess, attest, or certify compliance.
Get started
Run the seven against your own project.
Connect Supabase and GitHub, then read the verdict alongside the evidence that produced it.