SecureVibe

Prevention-first security for AI-written code. SecureVibe feeds your AI coding assistant signed SKILL.md security knowledge at the point of code generation, then backs it with deterministic scanners and a CI gate. One static Go binary — secure-vibe — is both the CLI/gate and the MCP server. Offline, keyless, no telemetry, Ed25519-signed releases.

Maintained by ShieldNet360 · MIT — free to fork, embed, and ship commercially.

Install

npm bundles the library data (skills + vuln DB), so it works on macOS, Linux, and Windows out of the box:

npx -y @shieldnet360/secure-vibe <command>   # run on demand: audit · check · init · mcp · status
npm install -g @shieldnet360/secure-vibe      # …or install globally for a persistent `secure-vibe`

Downloads the binary and library data from the latest GitHub release, verifies SHA-256, and installs to a per-user dir (~/.local/share/secure-vibe / %LocalAppData%\secure-vibe). Override with SECURE_VIBE_{BIN,DATA}_DIR.

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/ShieldNet-360/secure-vibe/main/install.sh | sh

# Windows (PowerShell)
irm https://raw.githubusercontent.com/ShieldNet-360/secure-vibe/main/install.ps1 | iex

go install github.com/shieldnet-360/secure-vibe/cmd/secure-vibe@latest builds only the binary (no bundled data). Point it at a library checkout via --path / $SECURE_VIBE_LIBRARY_PATH.

Quick start

# 1 · embed security skills into your assistant (writes CLAUDE.md / .cursorrules / …)
secure-vibe init --tool claude        # claude · cursor · copilot · codex · windsurf · cline · devin

# 2 · audit code — auto-detects secrets / bad deps / Dockerfile / GitHub Actions,
#     deduped + ranked + fixtures triaged. Pass files, a dir, or --diff for a PR.
secure-vibe audit .

# 3 · gate a build (CI / pre-commit) — non-zero exit on findings, SARIF for Code Scanning
secure-vibe audit . --fail-on high --format sarif

# 4 · vet one package before adding it
secure-vibe check [email protected] -e npm

# 5 · the MCP server (20 tools over stdio) — this is the command clients spawn:
secure-vibe mcp                                    # or: npx -y @shieldnet360/secure-vibe mcp
#    register it with Claude Code in one line:
claude mcp add SecureVibe -- npx -y @shieldnet360/secure-vibe mcp
#    Cursor · Windsurf · VS Code · Cline · Zed → docs/reference/mcp-clients.md

See it catch something

$ secure-vibe check [email protected] -e npm
=== check [email protected] (npm) ===
Malicious entries:  1
  ! MALICIOUS  [critical]  event-stream — Maintainer account compromised; malicious
    flatmap-stream dependency added to steal cryptocurrency wallets

$ secure-vibe audit Dockerfile --fail-on high   # FROM ubuntu:latest / USER root
=== secure-vibe audit: Dockerfile ===
Findings: 3   (critical: 1, high: 2)
$ echo $?
1

Use in CI

Audit a PR's diff, fail the build on findings, upload to Code Scanning, and comment on the PR — one step:

# .github/workflows/secure-vibe.yml
on: [pull_request]
permissions: { contents: read, security-events: write, pull-requests: write }
jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: ShieldNet-360/secure-vibe@v1
        with:
          diff: origin/${{ github.base_ref }}   # PR-scoped; omit to audit the whole tree
          fail-on: high

Any other CI (or a pre-commit hook): npx -y @shieldnet360/secure-vibe audit . --fail-on high — add --diff origin/main for PR scope, --format sarif for Code Scanning.

What you get

The lifecycle is PREVENT → DETECT → ENFORCE → VERIFY → LEARN. The binary is deterministic and passive — it never calls an LLM and never sends attack traffic; the reasoning and dynamic verification are your coding agent's job, so the tool stays keyless and works the same in Claude Code, Codex, Gemini, or plain CI.

• PREVENT — 33 signed skills across 8 assistants, consulted as code is written.
• DETECT — one command, audit: fans the deterministic scanners (secrets, dependencies via a curated malicious / typosquat DB + CVE / OSV across 10 ecosystems, Dockerfile, GitHub Actions) across the tree (or a PR's --diff), then dedups, ranks by severity, and triages likely fixtures.
• ENFORCE — audit --fail-on <severity> exits non-zero for CI; --format sarif for Code Scanning, --report-dir for an HTML + PDF report, --no-triage for a strict gate.
• VERIFY — the coding agent confirms candidates dynamically using two scope-gated MCP primitives — http_probe (send one crafted request) and oob_listener (catch blind callbacks) — guided by the dynamic-verification skill. In CI (no agent), verification is a committed regression test (security-regression-tests skill), not a live probe. The binary never sends attack traffic on its own.
• LEARN — two feedback loops that improve the tool from real use. Data: secure-vibe contribute add -p <pkg> -e npm writes a .secure-vibe/overlay.json the gate enforces — commit it (team) or point $SECURE_VIBE_OVERLAY at a shared file (org). Knowledge: when the agent finds a skill is wrong or missing a fact, it records an inert proposal via the propose_skill_update MCP tool; review with secure-vibe contribute skill, then edit the skill and re-sign. Proposals never touch the signed skill on their own — a human review is the gate.

Narrow by design. Detection is four deterministic scanners, not a general SAST. It catches known patterns and known-bad packages with near-zero false positives; it does not claim to find every vulnerability. The semantic and dynamic depth comes from the agent driving it.

Commands

Full reference: docs/reference/cli.md · secure-vibe --help.

Docs

• ARCHITECTURE.md — design, compiler, update protocol, repo layout.
• docs/ — guides (developer · devops · security · evaluator), install, air-gapped, team rollout.
• docs/reference/mcp-clients.md — connect the MCP server to any agent (Claude Code · Cursor · Windsurf · VS Code · Cline · Zed) · docs/reference/mcp-tools.md — the 20 MCP tools · skills/ — the 33-skill catalogue.
• SIGNING.md — Ed25519 release signing · CONTRIBUTING.md · SECURITY.md.

Platform support

License

MIT — Copyright (c) 2024-2026 ShieldNet360. Free to fork, embed, and ship commercially; please preserve the notice and attribution.