Pentect is a local security boundary for AI coding tools. It replaces secrets and sensitive data with opaque handles before requests reach a model provider, then resolves those handles only at trusted local tool boundaries.

The problem with redaction

Conventional redaction protects a value by removing its meaning:

DATABASE_URL=[REDACTED]

The model can no longer use that value in a command. Pentect preserves a typed, stable reference instead:

DATABASE_URL=<<DATABASE_URL_4ce8a3b0a6f64e12>>

The model can copy the handle into a completed tool call. Pentect resolves it immediately before the trusted local client executes that call. The provider never needs the plaintext.

What Pentect is not

• It is not a replacement chat application.
• It is not a password manager or secret vault.
• It does not grant tools permission to use a secret.
• It cannot guarantee detection of every future or unknown sensitive format.

Pentect reduces exposure at the provider boundary. Existing client permissions, sandboxing, account controls, and least-privilege credentials still matter.

Supported surfaces

Pentect currently integrates with Codex CLI, Claude Code, Codex App, and supported Claude Desktop routes. It also provides standalone masking and execution commands, custom upstream routing, and sandboxed plugins.

See Compatibility for the release-tested matrix.