StingWard MCP server
Commit Nov 5, 2026Audits AI-built applications, Claude Code skills, and MCP servers to identify security vulnerabilities, leaked secrets, and malicious dependencies.
Also known as HiveControlAgents/StingWard
- Stars
- 3
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 5, 2026
- License
- Apache-2.0
What is StingWard MCP?
StingWard audits AI-built applications, Claude Code skills, and MCP servers to identify security vulnerabilities, leaked secrets, and malicious dependencies.
Summary generated from the project's README
About StingWard MCP server
Securing AI-generated applications and agentic tools is what StingWard enables for developers and security researchers. It identifies vulnerabilities, leaked secrets, and malicious dependencies within AI-built apps, Claude Code skills, and MCP servers. The tool works locally via terminal commands or through integrations like Claude Code, Codex, and GitHub Actions. It can audit a local project folder, a specific skill or plugin, or a remote repository directly from GitHub.
Users can run audits using pipx, uvx, or npx, and the system operates with zero dependencies using pure Python 3.9+. It checks lockfiles against OSV.dev and ranks vulnerabilities using CISA KEV and EPSS data. The server detects a wide range of issues, including invisible Unicode instructions in prompts, exposed API keys like OpenAI or AWS, and unsafe patterns like model answers reaching a shell or eval. It provides a final grade from A to F along with a remediation plan to help close security holes.
What you can do with it
- 1Audit a third-party skillVerify if a new MCP server or Claude Code skill is safe to install by checking for hidden instructions or malicious scripts.
- 2Scan for leaked credentialsDetect if sensitive API keys or private credentials have been accidentally committed to your codebase.
- 3Check for malicious dependenciesIdentify known CVEs and malicious packages within your project's lockfiles across various package managers.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
STINGWARD_ONLINE
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablehivecontrolagents/stingward is public and reachablegithub.com/hivecontrolagents/stingward ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-05github.com/hivecontrolagents/stingward ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the modelInstruction aimed at the model found in the description or READMEFinding
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubLots of people now ship apps, Claude Code skills and MCP servers written with AI. Few of them know what a CVE is, why NEXT_PUBLIC_OPENAI_API_KEY is a problem, or that a skill can hide instructions only the model can read. Stingward finds those holes and tells you, in plain words, how to close each one.
Quick start
| Where | How |
|---|---|
| Claude Code | /plugin marketplace add HiveControlAgents/StingWard then /plugin install stingward@stingward, and ask "is my project secure?" or "can I install github.com/x/y?" |
| Codex | codex plugin marketplace add HiveControlAgents/StingWard then codex plugin add stingward@stingward, and ask the same questions. Codex asks you before the dependency check goes online |
| Terminal | pipx install stingward (or uvx stingward, or npx stingward with Python 3.9+ installed), then run stingward in your project |
| GitHub | Add the Action: it comments on every PR, fills the Security tab and re-checks daily |
stingward # audit the project in the current folder
stingward skill ./some-skill # audit a skill, plugin or MCP server before installing it
stingward skill owner/repo # ...straight from GitHub (also accepts .../tree/branch/subfolder URLs)
stingward agents # audit the Claude Code, Claude Desktop, Cursor, Codex and Gemini setup on this machine
No account, no telemetry, no dependencies: pure Python 3.9+ on Linux, macOS and Windows.
What it finds
| Area | Examples |
|---|---|
| Dependencies | Known CVEs and malicious packages in npm, Yarn, pnpm, Bun, pip, Poetry, uv, Pipenv, PDM, Go, Cargo, Bundler, Composer and NuGet lockfiles. Ranked by CISA KEV (exploited right now) and EPSS (likelihood of exploitation), with the single upgrade that fixes everything known. Without a lockfile, the lowest version each range allows (Pillow>=10, ^1.6.0) is checked too. So is the PHP or Node.js version in a Dockerfile, docker-compose.yml, .php-version, .nvmrc or Replit's .replit. |
| Secrets | API keys (OpenAI, Anthropic, GitHub, AWS, Google, Stripe, Slack...), private keys, Supabase service_role keys, webhooks, passwords in URLs, .env committed or missing from .gitignore. Values are never printed. |
| AI-built apps | Secrets behind NEXT_PUBLIC_ / VITE_ / EXPO_PUBLIC_, dangerouslyAllowBrowser, open CORS, Supabase tables without RLS, with using (true) or with policies that only check that someone is signed in, Edge Functions with the service_role key that let the visitor pick the record, database functions that run any SQL they receive, open Firebase rules, debug mode, JWT without signature check, crypto wallet keys and recovery phrases in the code, EIP-7702 authorizations that hand the wallet to a contract the visitor or the model picks. In code that calls a model: its answer reaching eval, a shell, SQL or HTML, request input inside the system prompt, AI routes with no login and no rate limit (with an SDK or with fetch to the provider, such as the Lovable AI Gateway), trust_remote_code and unsafe torch.load, agents with a Python REPL or shell tool. In MCP servers and agent tools: arguments the model writes reaching a shell, eval, SQL, a file path or a URL the server fetches. |
| Code | SQL, shell and code injection, pickle and unsafe YAML, XSS (innerHTML, dangerouslySetInnerHTML, v-html), TLS checks turned off, path traversal, SSRF, open redirects, Math.random for tokens and fast hashes for passwords. Python is read with ast and JavaScript/TypeScript with tight patterns that follow request data inside each function. |
| Agents and MCP | bypassPermissions, Bash(*), hooks that curl | sh, enableAllProjectMcpServers, MCP servers without a pinned version, tokens in .mcp.json, MCP over plain HTTP, vulnerable MCP packages such as mcp-remote (CVE-2025-6514). |
| Skills and prompts | Invisible Unicode (tag characters, Trojan Source, zero-width), "ignore previous instructions", "don't tell the user", reading ~/.ssh or ~/.aws and calling the internet, curl | bash, reverse shells, obfuscated code, capture hosts (webhook.site, ngrok...), install scripts, allowed-tools: Bash, tool poisoning in MCP tool descriptions. |
All 57 rules, with severity, examples and fixes: docs/rules.md.
How it works
- Lists your files, skipping dependency and build folders such as
node_modulesanddist. When auditing a third-party skill it does read build output, since that is what actually runs. - Reads lockfiles and asks OSV.dev about every package and version in one batch. CVEs in CISA KEV become critical; EPSS and NVD fill in likelihood and missing scores.
- Runs pattern checks for secrets, app mistakes, agent settings and skill content, aware of context: test folders, docs and examples count less, and in third-party skills reassuring wording can lower a finding to medium but never hide it.
- Writes the report in English or Portuguese: a verdict, what to fix first, findings worst first with where, why and how to fix (the same problem in several files is listed once, with every place), a grade from A to F and a badge.
--html FILEalso writes a single-file report to hand over or print to PDF: summary, remediation plan with rough effort, findings with the code excerpt (secrets masked), and scope and method.
Only package names and versions leave your machine (to api.osv.dev, www.cisa.gov, api.first.org and services.nvd.nist.gov). Your code never does. Responses are cached for a few hours in ~/.cache/stingward.
Stingward reads hostile content by design, so it is careful with it: git can never discover a repository inside audited files, the launchers never import code from the audited folder, and secrets, terminal escapes and markdown are neutralized in every report format. See SECURITY.md.
GitHub Action
# .github/workflows/stingward.yml
name: stingward
on:
pull_request:
push: { branches: [main] }
schedule: [{ cron: "0 9 * * *" }] # new CVEs show up for code that did not change
permissions: { contents: read, pull-requests: write, security-events: write }
jobs:
stingward:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: HiveControlAgents/StingWard@v0
with: { lang: en, fail-on: high }
It comments the report on the PR (and updates the same comment on every run), uploads SARIF to the repository's Security tab, writes the job summary and exposes the grade as steps.<id>.outputs.grade.
| Input | Default | |
|---|---|---|
path |
. |
Folder to audit |
fail-on |
high |
Fail the job at this severity or above: critical, high, medium, low, never |
lang |
en |
Report language: en or pt |
comment |
true |
Comment on pull requests (needs pull-requests: write) |
sarif |
true |
Upload SARIF to code scanning (needs security-events: write, plus actions: read in private repositories) |
category |
stingward |
Code scanning category for the SARIF. Give each run its own when one workflow audits several folders |
Command line reference
| Option | |
|---|---|
--lang en|pt |
Report language (default: STINGWARD_LANG, then your system locale) |
--format text|json|markdown|sarif|html |
What goes to standard output |
--json FILE, --markdown FILE, --sarif FILE, --html FILE |
Also write these files, in one run |
--fail-on critical|high|medium|low|never |
Exit code 1 at this severity or above (default high) |
--only deps,secrets,app,code,agents |
Run only some checks |
--offline |
Skip the online databases (dependencies are not checked) |
--no-ignore |
Do not apply the project's .stingwardignore or stingward:ignore comments, to audit someone else's project |
--no-color |
Plain text output |
--review-template FILE |
Write the human review file: one entry per finding, keeping what was already reviewed |
--review FILE |
Apply the human review: a status and a comment for each finding |
--client NAME, --notes FILE |
"Prepared for" on the cover, and the analyst summary (paragraphs, - lists, **bold**, `code`) |
| Exit code | Meaning |
|---|---|
0 |
Nothing at or above --fail-on |
1 |
Something at or above --fail-on was found |
2 |
Usage error: a wrong option, a folder that does not exist, a skill that could not be downloaded or an output file that could not be written |
3 |
OSV.dev did not answer, so dependencies were not checked, or answered only in part, so the check is incomplete. A CI gate should not pass silently; use --offline to skip them on purpose. |
4 |
No file was read, so there was nothing to grade: an empty folder, a wrong path, or in CI the code was not checked out first. A CI gate should not pass silently; --fail-on never accepts it. |
Environment: STINGWARD_LANG sets the default language, NVD_API_KEY raises the NVD rate limit and STINGWARD_CACHE moves the cache folder.
Grade: A means only low findings or none · B medium, no high · C one or two high · D three or more high · F any critical. Every report ends with a badge for your README:
[](https://github.com/HiveControlAgents/StingWard)
Human review: for an audit you hand over to someone, write the review file with stingward . --review-template review.json. Each entry points at a finding by its fingerprint (also in the JSON and SARIF output), which does not change when other parts of the file are edited. It is built from the flagged line with the secret already masked, so nothing derived from a secret goes into the review file. That is on purpose, and it means a key replaced by another key on the same line keeps its status: when a key is rotated, check that line again before you reuse the review. Set status to confirmed, false_positive, accepted_risk or fixed, add a comment, then run stingward . --review review.json --client "Client name" --notes notes.md --html report.html. False positives move to an appendix; they and the findings fixed during the audit leave the grade, the fix-first list and --fail-on. Accepted risks stay in the grade but leave the fix-first list. The text and Markdown reports follow the review too. Comments and notes are escaped and go through the same secret masking as code excerpts. For a review done by an AI, add "review_kind": "ai" to the review file: the reports then say automatic review instead of human review, accepted_risk is refused (only the client can accept a risk), and fixed is only accepted for a finding that is gone from a scan that could see it: not hidden by .stingwardignore, a stingward:ignore comment or the size and file limits.
False positives: add a line to .stingwardignore in the project root (RULE path/glob, RULE or path/glob), or put stingward:ignore in a comment on the line. When you audit someone else's project, --no-ignore turns both off. If a rule is wrong, tell us with a minimal sample.
Standards: findings carry CWE, OWASP Top 10 for LLM Applications 2026 (LLM01:2026...), OWASP Agentic Skills Top 10 (AST01..., v1 draft) and OWASP Top 10 for Agentic Applications 2026 (ASI01...) IDs, in the text report, JSON and SARIF tags.
Quality
- 400+ automated tests, plus online tests that query the real OSV, KEV, EPSS and NVD services across six ecosystems. Line coverage is above 90%.
- Every push runs 24 CI jobs: tests on Linux, macOS and Windows with Python 3.9 to 3.14, the PyPI and npm packages built and installed from scratch on all three systems, the Claude Code plugin validated, and the Action run against sample projects.
- Stingward audits itself in CI and must stay free of high and critical findings; its own grade is the A badge above.
- False positives are measured against popular public skill collections and MCP servers: no critical or high findings that are not real, and only a handful of medium ones per repository. Each fixed false positive becomes a regression test.
- Security review of Stingward itself, for example making sure a skill folder crafted as a git repository cannot run commands, and that a secret can never reach a report. Found issues are fixed with a test that fails on the old code.
python -m unittest discover -s tests -t . # offline, a few seconds
STINGWARD_ONLINE=1 python -m unittest discover -s tests -t . # also hits the real databases
FAQ
Does a clean report mean my project is safe? No. It means the common mistakes are not there. Stingward is a fast first line, not a penetration test.
Does my code leave my machine? No. Only package names and versions are sent, to look up vulnerabilities.
How is it different from npm audit, Dependabot or a secret scanner? Those each cover one piece. Stingward puts dependencies, secrets and the mistakes specific to AI-built apps, agents, skills and MCP servers in one report, written for the person who built the project, and can audit a skill before you install it.
Why does npx stingward need Python? The engine is a single pure-Python package with no dependencies; the npm package is a thin launcher so JavaScript users can run it without pip. On Windows, npx does not find the .bat shims of pyenv-win: install Python from python.org, or use pip install stingward.
Does it work in Codex or other agents? Yes. In Codex, install the plugin as in the Quick start. Codex blocks the network by default, so it asks you once before the dependency check reaches OSV.dev; say yes, or the dependencies go unchecked. For another agent that reads SKILL.md skills, copy the skills/stingward folder into its skills folder and install the command line with pipx install stingward, which the skill uses when the engine is not next to it.
Can it check a private repository or a skill on my disk? Yes. Point it at a folder, or at owner/repo with your usual git credentials.
Scope
Stingward reports to whoever runs it, about code they chose to check. It does not crawl other people's repositories or open issues on them.
Contributing
Bug reports, false positives and new rules are welcome. Read CONTRIBUTING.md to get set up, and report security problems in Stingward itself privately as described in SECURITY.md.
Vulnerability data: OSV.dev (aggregating GitHub Advisories, NVD, PyPA, RustSec, Go and OpenSSF malicious packages), CISA KEV, FIRST EPSS and NVD.
Apache 2.0 © Isaque Moutta
StingWard MCP questions
What does StingWard find?
It identifies known CVEs, malicious packages, exposed API keys, secrets in code, vulnerabilities in AI-built apps, and security risks in MCP servers and agent tools.
How can I use StingWard with Claude Code?
You can add it via the plugin marketplace using /plugin marketplace add HiveControlAgents/StingWard and then install it with /plugin install stingward@stingward.
Does StingWard require an account?
No, it operates without an account, telemetry, or dependencies.