secretgate

name	secretgate
description	Use when the user wants to protect secrets/credentials from being sent to an LLM by a coding agent, install or manage a local secrets firewall, or scan text/files/repos for leaked credentials. secretgate detects secrets (221 gitleaks-derived rules + entropy + Luhn, zero-dependency engine, `node scripts/secretgate.mjs`) in prompts, file reads and tool output BEFORE they reach the LLM API, redacts them to stable placeholders (SECRETGATE_xxx) and restores the real values locally when the agent writes them back to files. Ships installers for Claude Code (hooks), OpenAI Codex CLI (hooks) and OpenCode (plugin). Triggers: 'protect my secrets', 'secrets firewall', 'stop sending credentials to the LLM', 'install secretgate', 'scan for secrets', 'redact secrets before sending', 'block credentials in prompts'.
license	MIT
metadata	{"version":"1.1.6"}

secretgate — local secrets firewall for coding agents

A credential must never travel to an LLM API. secretgate enforces that locally, with a deterministic engine (node scripts/secretgate.mjs, no npm install, no keys): 221 rules converted from gitleaks + entropy gates + Luhn-checked cards. Where the agent's hook API allows rewriting, secrets are redacted to stable placeholders and restored locally when the agent writes them back (redact-and-restore); where it only allows blocking (prompts on Claude Code/Codex), the prompt is blocked with a redacted copy to resend.

Activation → install (do this first when the skill triggers)

When this skill activates (the user asks to protect secrets / install secretgate / set up the firewall), run the doctor, then wire whatever is not yet wired. secretgate is a skill, not an npm package — you run its own bundle:

node scripts/secretgate.mjs status                  # what is already wired
node scripts/secretgate.mjs install --claude-code   # hooks + permissions.deny in ~/.claude/settings.json
node scripts/secretgate.mjs install --codex         # ~/.codex/hooks.json + [features] hooks = true
node scripts/secretgate.mjs install --opencode      # self-contained plugin in ~/.config/opencode/plugin/
node scripts/secretgate.mjs install --all           # all three

--project scopes the Claude Code install to ./.claude/settings.json. install pins a copy of the bundle under ~/.secretgate/bin/ and wires the agent to call THAT, so it keeps working after the skill cache is evicted. Always tell the user to RESTART the agent session afterwards — hooks load at startup. Once wired, redaction is automatic and deterministic (the hook does it, not you).

Commands

Command	What it does
status	Doctor: per-agent wiring, engine, vault health, limitations
scan <file|dir|->	Scan for secrets (exit 1 on findings; --json; never prints raw secrets)
pipe	stdin -> stdout with secrets redacted to placeholders
allow <value> / allow --rule <id> / allow --path <glob>	Allowlist (values stored as SHA-256, never in clear)
vault list / vault clear	Inspect placeholder mappings (never shows secrets)
uninstall --<agent>	Remove exactly what install added

How the protection works (tell the user when asked)

• Prompts: Claude Code/Codex → blocked with a redacted copy (add [allow-secret] to bypass once). OpenCode → silently redacted in place.
• File reads: .env*, keys, ~/.aws, ~/.ssh… are denied outright (.env.example/.sample/.template stay readable). The value never enters the model.
• Tool/bash output: redacted to SECRETGATE_<hash> placeholders (Claude Code + OpenCode; Codex cannot rewrite output yet).
• Restore: when the agent writes a placeholder into a file (Write/Edit), the REAL value lands on disk. Bash restore is OFF by default (prompt-injection exfiltration guard) — enable with restoreBash: true in ~/.secretgate/config.json.
• Clean tool calls: the Claude Code pre-tool-use hook answers minimal JSON ({}, decision-free = normal permission flow) instead of empty stdout. Empty hook output triggers a Claude Code bug (anthropics/claude-code#77782) that forces an interactive permission prompt on every matched tool call, even in auto mode — and the documented "defer" decision value is rejected by Claude Code <= 2.1.212, so {} is the abstain that works everywhere.

NOT covered (be honest with the user)

• Claude Code @file mentions inline content without firing tool hooks (permissions.deny rules are the only cover there).
• Codex non-interactive runs (codex exec) — upstream bug, hooks don't fire.
• Images/screenshots/clipboard, MCP tool traffic, and anything already in the conversation before install.
• If a hook process is killed/times out, the agent proceeds (agent-side fail-open); secretgate itself fails CLOSED on its own errors for pre-events.
• A blocked prompt is echoed back locally by Claude Code (Original prompt: in the terminal and in claude -p JSON result) — the API never sees it, but treat piped headless output as sensitive.

False positives

Prefer the narrowest fix: inline # pragma: allowlist secret (or gitleaks:allow) on the line → allow <value> (hashed) → allow --path <glob> → allow --rule <id> (last resort). A project can commit extra allowlist entries in .secretgate.json ({"allowlist": {"paths": […]}}).