Sentrail documentation
Security and data access
Updated August 17, 2026
Access follows your connected project
Sentrail uses connected GitHub, Supabase and Vercel accounts to inspect project evidence. Connect only projects you intend to review. Provider permissions determine what those connections can access.
MCP tokens are scoped to a Sentrail project, sent in the Authorization header and stored as hashes. Revoke tokens you no longer use. Tool calls are recorded in the project’s MCP audit log.
Read-only does not mean anonymous
The database query tool accepts queries starting with SELECT or WITH and asks the Supabase Management API to execute them with read_only enabled. The provider’s read-only execution is the enforcement boundary; checking the first SQL keyword alone is not a security guarantee.
The tool returns up to 100 result rows to the assistant or connected MCP client. This response limit is not anonymization and does not limit which rows the database may examine.
Review a proposed change before applying it
Database writes are prepared as pending actions for explicit approval. A preview is not an applied change. Inspect the target, SQL and expected effect, then check the execution receipt and verification result.
Read the remediation workflow for the approval and verification steps.
Repository content is untrusted input
Repository snippets and selected finding fields pass through length limits, known secret-pattern filtering and filtering of common prompt-injection markers before model use. These pattern-based controls do not detect every secret or malicious instruction, and they are not a guarantee against prompt injection or data disclosure.
Project scope, token revocation and approval requirements provide additional boundaries. Review agent requests and proposed changes; do not put credentials or unnecessary personal data into prompts. Data returned to an external coding client is also subject to that client’s configuration and provider policies.
Evidence and assurance
A Sentrail verdict describes the available evidence within the reviewed scope. It is not a security certification or a guarantee that an application is safe. Missing or incomplete checks must be reviewed as such.
See our privacy policy, subprocessors and responsible disclosure policy. Contact us before connecting sensitive workloads if you need an assurance document that is not published.