[](https://www.npmjs.com/package/@zfuzz/mcp)
[](LICENSE)
[](package.json)
**[Install](#-add-it-in-10-seconds) · [Tools](#whats-under-the-hood) · [Agents](#agent-compatibility-matrix) · [When to use](#when-to-use--when-to-skip) · [Security](SECURITY.md) · [zfuzz.com](https://zfuzz.com)**
# Zfuzz — security for everyone who builds with AI
**You ask. Your AI answers with real scanners — not guesses.**
Catch SQL injection, leaked API keys, and vulnerable dependencies **while you build** — no security background required. You don't run tools. You don't read dashboards. You just talk to your AI, and Zfuzz gives it real answers.
`$0` · **Apache-2.0** · No account · No API key · 100% local · Nothing leaves your machine
Made for **vibe coders and developers alike.** It plugs straight into the AI tools you already use — Cursor, Claude Code, Codex, Gemini CLI, OpenCode — and the web builders AI Studio, v0, and Lovable.
---
## ⚡ Add it in 10 seconds
**Pick your tool. Copy one line. Done.** No setup wizard, no account, no config to learn.
### Cursor — one click
[](cursor://anysphere.cursor-deeplink/mcp/install?name=zfuzz&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsIkB6ZnV6ei9tY3AiXX0=)
Click the button → Cursor opens → press **Install**. That's your two clicks.
### Claude Code
Paste this in the chat and hit enter:
```bash
claude mcp add zfuzz -- npx -y @zfuzz/mcp
```
### VS Code (Copilot)
```bash
code --add-mcp '{"name":"zfuzz","command":"npx","args":["-y","@zfuzz/mcp"]}'
```
### Codex
```bash
codex mcp add zfuzz npx -y @zfuzz/mcp
```
### Gemini CLI
```bash
gemini mcp add zfuzz npx -y @zfuzz/mcp
```
### Lovable · v0 · AI Studio — in your browser, no install
Open your tool's **MCP / Integrations** settings and paste this URL:
```
https://api.zfuzz.com/mcp
```
Other tools — OpenCode, Claude Desktop, or a manual config file
**Claude Desktop** — Settings → Developer → Edit Config, then add:
```json
{
"mcpServers": {
"zfuzz": { "command": "npx", "args": ["-y", "@zfuzz/mcp"] }
}
}
```
**OpenCode** — add to `opencode.json`:
```json
{
"mcp": {
"zfuzz": { "type": "local", "command": ["npx", "-y", "@zfuzz/mcp"] }
}
}
```
**Self-host the web (HTTP) endpoint** instead of using the hosted URL:
```bash
npx -y @zfuzz/mcp --transport http --port 8099
```
Then point your tool at `http://localhost:8099/mcp`.
_(Use port 8099, not 8090 — 8090 is the Zfuzz platform's own port.)_
That's it. Keep talking to your AI like normal — the first time your code needs a security check, it calls Zfuzz on its own.
---
## Agent compatibility matrix
Works with every major AI coding agent — and any MCP-compatible client. One line locally, or a hosted URL in the browser. Same 10 tools everywhere.
| Agent | Zfuzz | One-line setup |
|---|:---:|---|
| **Claude Code** | ✅ | `claude mcp add zfuzz -- npx -y @zfuzz/mcp` |
| **Cursor** | ✅ | one-click button above (or `.cursor/mcp.json`) |
| **VS Code · Copilot** | ✅ | `code --add-mcp '{"name":"zfuzz",…}'` |
| **Codex** | ✅ | `codex mcp add zfuzz npx -y @zfuzz/mcp` |
| **Gemini CLI** | ✅ | `gemini mcp add zfuzz npx -y @zfuzz/mcp` |
| **Claude Desktop** | ✅ | add `zfuzz` to the MCP config |
| **OpenCode** | ✅ | add `zfuzz` to `opencode.json` |
| **Lovable · v0 · AI Studio** | ✅ | paste the hosted URL `https://api.zfuzz.com/mcp` |
| **Any other MCP client** | ✅ | stdio `npx -y @zfuzz/mcp`, or the hosted URL in-browser |
_No per-agent config to learn — the exact copy-paste line for each is under [Add it in 10 seconds](#-add-it-in-10-seconds)._
---
## What it feels like
> **You:** Is my login page safe?
>
> **AI:** _[scans your file]_ Found 2 issues:
> 1. **SQL injection** on line 42 — anyone could read your whole database.
> 2. **No rate limiting** on the login — bots can guess passwords forever.
> Here's the fix… _[writes the patch]_
> **You:** Did I leave any passwords or keys in my code?
>
> **AI:** _[scans the project]_ Yes — your **OpenAI key** is hard-coded in `config.js`. Anyone who sees this repo can use it on your bill. Move it to an environment variable — here's how…
> **You:** Are the packages I installed safe?
>
> **AI:** _[checks your dependencies]_ 2 have known security holes. `lodash` lets attackers run code on your server. One command fixes both: `npm update lodash axios`.
You never typed a command, opened a scanner, or read a report. You just asked.
---
## See it in action
_You vibecode it. It ships with holes you can't see. Zfuzz catches them all — before you ship, for $0._
---
## You don't need to understand security
- You **never** run a scanner yourself — your AI does, automatically, when it matters.
- You **never** read a dashboard — answers come back in plain English, in your chat.
- You **never** pay and **nothing leaves your computer** — no account, no cloud, no API key.
**If you can copy-paste one line, you're covered.**
---
## What's under the hood
10 real tools your AI can call (it picks the right one — you don't have to):
| Tool | In plain English |
|------|------------------|
| `scan_code` | Finds bugs attackers exploit — 441 rules, 7 languages (Python, JS/TS, Go, Java, Rust, Ruby, PHP). |
| `scan_secrets` | Catches leaked passwords & API keys — 419 patterns (AWS, GitHub, Stripe, OpenAI, Anthropic…). |
| `scan_dependencies` | Flags packages with known security holes (CVEs via OSV.dev). |
| `scan_mcp_config` | Audits the MCP servers you install — hidden instructions, unicode tricks, wildcard permissions, unauthenticated remote servers. Returns a risk score + install verdict. |
| `scan_skill` | Vets an agent skill (its `SKILL.md` + bundled scripts) before you install it — hidden instructions, unicode deception, over-broad triggers, wildcard tool grants, and dangerous code or leaked secrets in its scripts. |
| `reconcile_permissions` | Compares the tools your agents actually use against what they declared — flags capabilities used but never declared, and permissions granted but never used (least privilege, from runtime telemetry). |
| `check_mitre` | Maps any finding to real-world attack techniques (MITRE ATT&CK). |
| `threat_model` | Asks "how could this be attacked?" across your whole project (STRIDE + MITRE). |
| `explain_finding` | Explains any vulnerability — and the fix — in everyday language. |
| `search_security_procedures` | Looks up 754 security playbooks (incident response, hardening, compliance). |
Built in **Rust** for sub-second answers. The AI brain is your editor's own model (Claude/GPT) — **Zfuzz adds the security muscle, not another subscription.**
### What's inside
| | |
|---|---|
| **441** SAST rules across **7** languages | **419** secret-detection patterns |
| **75+** MITRE ATT&CK techniques | **754** security playbooks |
| **Rust** — sub-second scans | **100% local** — nothing leaves your machine |
### What a scan actually returns
> **You:** Is this skill safe to install?
>
> **AI:** _[runs `scan_skill`]_
> ```
> Verdict: DO NOT INSTALL (risk 100/100, 6 findings)
> CRIT ZF-C001 Prompt Injection Pattern Detected @ description
> MED ZF-SK001 Overly Broad Skill Trigger @ description
> HIGH ZF-LP001 Wildcard Permission Grant (Bash(*), *) @ allowed-tools
> CRIT ZF-C002 Pipe-to-Shell Download-Exec (curl | bash) @ setup.sh
> HIGH ZF-H001 Hardcoded Secret: AWS Access Key ID @ setup.sh
> ```
> This skill hides "ignore all previous instructions" in its description, grants
> itself every tool, and its setup script pipes a remote payload straight into a
> shell. Don't install it.
Every finding is from a real deterministic scanner — there is no model guessing in the scan path.
---
## Why not Snyk or Semgrep?
| | Snyk / Semgrep | Zfuzz |
|---|---|---|
| **Where** | A pipeline, 5–10 min after you push | Right in your editor, in seconds |
| **When** | After the bug shipped | While you're writing it |
| **How** | A dashboard + email alerts | A normal conversation with your AI |
| **For non-coders** | No — built for security teams | **Yes — built for you** |
| **Cost** | $25–100 / dev / month | **Free, forever** |
---
## When to use · When to skip
**Great fit if you…**
- Vibecode apps with AI (Lovable, Cursor, v0, Bolt, Claude Code) and want them checked **before you ship** — without learning security.
- Want leaked keys, injectable code, and risky dependencies caught right in your editor, in plain English.
- Want to vet an **MCP server or agent skill _before_ you install it**.
**Skip it if you…**
- Already run a full security team + pipeline and don't want findings in the editor.
- Work fully offline with no Node.js (the scanner binary is pulled via `npx` / `@zfuzz/cli`).
- Need a hosted dashboard with audit logs and SSO — that's the [Zfuzz platform](https://zfuzz.com), not the free MCP.
---
## Updating
`npx -y @zfuzz/mcp` always resolves the latest published version — most people never do anything.
```bash
npm view @zfuzz/mcp version # see the latest release
npm i -g @zfuzz/cli@latest # only if you pinned the CLI globally
```
Pinned a version in your MCP config? Bump it (or drop the pin) and restart your agent.
---
## Documentation
| Start here | Go deeper |
|---|---|
| [User guide](USER_GUIDE.md) — install & first scan | [What it scans](PRODUCT.md) — coverage & approach |
| [Security policy](SECURITY.md) — reporting & local-first design | [Contributing](CONTRIBUTING.md) — dev setup |
| [Add it in 10 seconds](#-add-it-in-10-seconds) — every agent | [zfuzz.com](https://zfuzz.com) — the full platform |
---
## Free · Open · Local
No API keys. No cloud account. No telemetry. Runs 100% on your machine — your code never leaves it. Apache-2.0 licensed, open source.
---
## Contributing
Issues and PRs welcome — see [CONTRIBUTING.md](CONTRIBUTING.md) for dev setup. Found a vulnerability? Please follow [SECURITY.md](SECURITY.md) — don't open a public issue.
---
## Community & support
- **Questions / bugs** — [open an issue](../../issues) on this repo.
- **Security reports** — see [SECURITY.md](SECURITY.md).
- **The platform** — [zfuzz.com](https://zfuzz.com).
---
## License
[Apache-2.0](LICENSE) — free & open source. © Zfuzz
Part of the [Zfuzz](https://zfuzz.com) security platform.