Hear every voice in your language
실시간 AI 더빙과 양언어 자막을 제공, 무료·오픈소스·프라이버시 보장
Claude Code용 AI 기반 보안 스캔 MCP 플러그인
AI-powered security scanning MCP plugin for Claude Code
플러그인Python · ShellGitHub @ArmisSecurity직접 설치운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
메이커가 쓴 소개는 위의 한 줄이 전부입니다. 더 자세한 내용은 저장소 README에서 읽을 수 있습니다. 저장소 README ↗
README 전체를 불러오는 중입니다.
README 전문 보기 ↗AI-powered security scanning for Claude Code, Cursor, VS Code (GitHub Copilot), Gemini CLI, GitHub Copilot CLI, Codex CLI, and Cline. Scans code, files, and git diffs for vulnerabilities in real-time using the Armis scanning API.
scan_code — Scan a code snippet for vulnerabilitiesscan_file — Scan a file on diskscan_files — Scan up to 20 files in one call (one approval prompt in editors that confirm each tool call)scan_diff — Scan git changes (staged, unstaged, or diff against a branch)approve_findings — Approve findings after user consent (for shipping with known risks)debug_config — Check scanner configuration status/security-scan — Claude Code slash command for on-demand scanning# 1. Clone the repository
git clone https://github.com/ArmisSecurity/armis-appsec-mcp.git
cd armis-appsec-mcp
# 2. Create credentials
cat > .env << 'EOF'
ARMIS_CLIENT_ID=<your-client-id>
ARMIS_CLIENT_SECRET=<your-client-secret>
EOF
chmod 600 .env
# 3. Generate config for your client
make setup CLIENT=cursor # or: vscode, gemini, copilot
Contact the Armis AppSec team if you don't have credentials.
Run make setup CLIENT=cursor and copy the output to ~/.cursor/mcp.json (user-level) or .cursor/mcp.json (workspace-level).
Or manually add to your config:
{
"mcpServers": {
"armis-scanner": {
"command": "/path/to/armis-appsec-mcp/run.sh",
"args": []
}
}
}
Run make setup CLIENT=vscode and copy the output to .vscode/mcp.json in your project.
Or manually add:
{
"servers": {
"armis-scanner": {
"type": "stdio",
"command": "/path/to/armis-appsec-mcp/run.sh",
"args": []
}
}
}
Enable MCP in VS Code settings if not already: github.copilot.chat.mcp.enabled: true.
Run make setup CLIENT=gemini and copy the output to ~/.gemini/settings.json (user-level) or .gemini/settings.json (project-level).
Or manually add the mcpServers block to your settings.json:
{
"mcpServers": {
"armis-scanner": {
"command": "/path/to/armis-appsec-mcp/run.sh",
"args": []
}
}
}
Run make setup CLIENT=copilot and copy the output to .mcp.json (workspace) or ~/.copilot/mcp-config.json (user-level).
Copilot CLI requires both command and args fields. A config without args will be ignored.
Add the MCP server to your Codex CLI config per its documentation. Then wire in the commit gate hook:
make setup CLIENT=codex # prints the hook config JSON
Merge the printed hooks block into your Codex CLI hooks config file (path varies by installation), replacing /absolute/path/to/armis-appsec-mcp with the actual clone path.
Add the MCP server via Cline's MCP settings panel. Then wire in the commit gate hook:
make setup CLIENT=cline # prints the hook config JSON
Merge the printed hooks block into your Cline settings.json, replacing /absolute/path/to/armis-appsec-mcp with the actual clone path.
Install via the plugin marketplace for the complete experience (hooks + slash command):
/plugin marketplace add ArmisSecurity/armis-appsec-mcp
/plugin install armis-appsec@armis-appsec-mcp
Then set credentials:
PLUGIN_DIR="$(ls -dt ~/.claude/plugins/cache/armis-appsec-mcp/armis-appsec/*/ | head -1)"
cat > "$PLUGIN_DIR/.env" << 'EOF'
ARMIS_CLIENT_ID=<your-client-id>
ARMIS_CLIENT_SECRET=<your-client-secret>
EOF
chmod 600 "$PLUGIN_DIR/.env"
| Feature | Claude Code | Cursor | VS Code | Gemini | Copilot CLI | Codex CLI | Cline |
|---|---|---|---|---|---|---|---|
| MCP tools (all 5) | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Commit gate (hard) | Native hook | Native hook | Git hook | Native hook | Native hook | Native hook | Native hook |
| Commit gate (soft) | Native hook | .cursor/rules | instructions | AGENTS.md | — | AGENTS.md | — |
| /security-scan | Yes | — | — | — | — | — | — |
"Native hook" = PreToolUse hook wired into the client's tool pipeline (blocks the command before it runs, injects a scan instruction). "Git hook" = portable pre-commit script (installed via make install-hooks). VS Code is the only client without a native hook template.
For a client-agnostic commit gate that works regardless of which AI tool you use:
make install-hooks
This installs a git pre-commit hook that verifies the scan-pass (stored inside .git/, so it never clutters your working tree) before allowing commits. It fails open by default (plugin bugs never block developers). Set APPSEC_HOOK_STRICT=1 for fail-closed behavior.
To remove: make uninstall-hooks
To test uncommitted changes from a clone end-to-end in Claude Code (or any client that loads the installed plugin), point the installed plugin at your working tree:
make dev-install # backs up the installed plugin, symlinks it -> this repo
# ...restart Claude Code, then test...
make dev-uninstall # restores the backed-up plugin exactly
make dev-status # show whether dev mode is active
dev-install backs up the real plugin to latest.bak, symlinks latest to this repo, and copies the installed .env (credentials) over so the launcher's preflight still passes. Restart Claude Code after each install/uninstall — MCP servers launch at session start. Override the cache location with PLUGIN_CACHE=... if your plugins live elsewhere.
/security-scan
Or ask your AI assistant: "scan staged changes for security issues"
/security-scan path/to/file.py
/security-scan ref=main
Paste code into the conversation and ask:
Is this code secure?
When the git pre-commit hook is installed, or when using Claude Code's native hooks:
If HIGH/CRITICAL findings are found, the assistant will attempt to fix them. If findings remain after remediation, it asks for your approval before proceeding.
| Environment Variable | Default | Description |
|---|---|---|
ARMIS_CLIENT_ID | (required) | Client ID for authentication |
ARMIS_CLIENT_SECRET | (required) | Client secret for authentication |
APPSEC_ENV | prod | dev or prod — selects API endpoint |
APPSEC_API_URL | (auto) | Override the API base URL |
APPSEC_SCAN_TIMEOUT | 180 | Seconds one scan may take (auth + API) before it fails with the step that stalled |
APPSEC_MAX_CONCURRENT_SCANS | 3 | Scan requests in flight at once (1–16); extra scans wait. Rate-limit (429) and 5xx responses are retried with backoff |
APPSEC_DEBUG | (unset) | Set to any value to enable debug logging |
APPSEC_TRANSPORT | stdio | MCP transport (stdio, sse) |
APPSEC_HOOK_STRICT | (unset) | Set to 1 for fail-closed git hook |
SSL_CERT_FILE / SSL_CERT_DIR / REQUESTS_CA_BUNDLE | (unset) | Explicit CA bundle; overrides the OS certificate store |
HTTPS_PROXY / ALL_PROXY / NO_PROXY | (unset) | Explicit proxy config; overrides the OS proxy settings |
Put ARMIS_CLIENT_ID / ARMIS_CLIENT_SECRET in the plugin's .env. Variables
exported in a terminal are not inherited by servers an editor launches
(VS Code started from the Start menu or Dock never sees them), and the server
then falls back to browser sign-in. Run debug_config to see which credentials
were found and from where, and check logs/server.log in the plugin directory
for the startup report.
.venv can add several seconds
to startup. If your editor gives up before the server is ready, ask IT to
exclude the plugin directory.APPSEC_SCAN_TIMEOUT fails with a message
naming the stage (obtaining a token, waiting for the API, or a pending browser
sign-in, including its URL and code). logs/server.log records each stage with
timings.The server trusts the OS certificate store (Windows cert store, macOS Keychain)
via truststore, so TLS-inspecting proxies such as Zscaler or Netskope work
without extra setup. CA precedence: SSL_CERT_FILE > SSL_CERT_DIR >
REQUESTS_CA_BUNDLE > OS store > certifi (used if truststore is unavailable).
With no proxy env vars set, the static system proxy (Windows Internet Options
registry, macOS System Settings) is used, honoring NO_PROXY and the OS bypass
list. PAC/WPAD auto-config scripts are not evaluated; on such networks set
HTTPS_PROXY explicitly.
The server logs to stderr and to <plugin dir>/logs/server.log (rotated at
~1 MB, 3 backups): startup config (version, Python, CA source, proxy, API URL),
each tool call with duration and outcome, and unhandled exceptions. Credentials,
tokens, and proxy passwords are never logged. The debug_config tool reports
the CA source, proxy, and log file path.
For teams that want a single shared scanner instance:
APPSEC_TRANSPORT=sse ./run.sh
Then configure clients to connect via HTTP instead of launching a local process.
macOS and Linux are fully supported. On native Windows, Claude Code launches the
plugin through run.cmd, which finds Git for Windows' bash.exe (Git for Windows
is already required by Claude Code) and runs run.sh. Set CLAUDE_CODE_GIT_BASH_PATH
if Git is installed somewhere unusual. WSL2 also works. For other MCP clients
(Cursor, VS Code, etc.), make setup launches natively via Git Bash — see
make setup CLIENT=....
make check # full CI gate (format + lint + typecheck + test)
make test # pytest only
pytest hooks/tests/test_pre_commit_scan.py -v # specific test file
+---------------------+
| Armis Cloud |
| POST /scan/fast |
+--------+------------+
^
| HTTPS (JWT Bearer)
+--------+------------+
| Scanner Core |
| scanner_core.py |
+--------+------------+
+-----+------+
| |
+-----v-----+ +---v---------+
| MCP Server | | Git Hook |
| server.py | | git-hooks/ |
+------------+ +-------------+
|
+------------+-------------+
| | |
Claude Cursor VS Code/
Code Gemini Copilot
Apache License 2.0 — see LICENSE for details.