PreHack MCP sunucusu

PreHack Connect

Connect your AI editor to PreHack, an autonomous AI pentester for your app. One short script, no API key.

PreHack connects to your GitHub, finds security flaws in your app, proves each one with a real attack in an isolated sandbox, and hands you the fix as a Pull Request. This repository is the small installer that plugs PreHack into the AI tool you already use (Claude Code or Cursor) through MCP.

What this repo is, and is not. It is only the connector: a skill plus one Python script. It contains no scanning engine, no security prompts, no attack skills and no secrets. All of that runs on PreHack's servers. See Security and trust.

Who is this for

• Developers who want to run and review security scans from inside Claude Code or Cursor.
• Founders and small teams shipping fast (often with AI-written code) with no security team and no budget for a manual pentest.
• Anyone who wants to know what an installer does to their machine before running it. This one is a single readable file, standard library only.

How it works

You (or your AI) run the connector. It never asks you to paste a key. It uses the OAuth 2.0 Device Authorization Grant (RFC 8628), the same kind of flow used by TVs and CLIs such as the GitHub CLI.

 Your machine                      PreHack server                 You, in a browser
 ------------                      --------------                 -----------------
 1. connector asks to start  --->  returns a short code + link
 2. prints the link and code  ------------------------------->   3. opens the link,
                                                                     signs in, picks free
                                                                     or subscribes, approves
 4. connector polls...        --->  "approved" + token   <-----
 5. writes the "prehack" entry into your editor's MCP config (file mode 0600)
 6. you restart the editor; PreHack's tools appear

1. The script asks the server for a one-time code and a verification link.
2. It prints them. You open the link in your own browser, create an account or sign in, then either Authorize (free) or Subscribe, and approve.
3. The script polls until approved (or until the code expires) and receives an MCP token.
4. It writes a single mcpServers.prehack entry into your editor's config.

Account creation, payment and approval always happen in your browser, never in the script and never on your behalf.

Quickstart

Requirements: Python 3.9+. Nothing to install.

Option A, ask your AI. In Claude Code or Cursor, paste:

install PreHack from github.com/Prehack-ai/prehack-connect

Your AI will fetch this repo, read the connector (one short Python file; that is the point) and run it with your OK. It will then show you the link and code to approve in your browser.

Option B, run it yourself:

git clone https://github.com/Prehack-ai/prehack-connect.git
cd prehack-connect

python3 scripts/prehack_connect.py                          # Claude Code, user scope
python3 scripts/prehack_connect.py --editor cursor
python3 scripts/prehack_connect.py --editor both --dry-run  # shows what it would do, writes nothing

Then restart your editor, list the server's tools (tools/list) and run postura_de_seguranca to confirm the connection.

Where it writes

Editor	--scope user (default)	--scope project
Claude Code	~/.claude.json	./.mcp.json
Cursor	~/.cursor/mcp.json	./.cursor/mcp.json

Safety behaviours of the script

• Idempotent: only touches the prehack key; if it is already identical, nothing changes.
• Backup first: before modifying an existing file it copies it to <file>.bak-YYYYMMDD-HHMMSS.
• Invalid JSON is never overwritten: such a file is left untouched.
• Token file permissions: the config is written with 0600.
• Dry run: --dry-run prints the plan and writes nothing.
• --scope project puts the token inside your repository. Do not commit it.

Options

Flag	Meaning
--editor claude-code|cursor|both	Which editor to configure (default claude-code)
--scope user|project	Global config or this project only (default user)
--transport http|sse	Streamable HTTP (default) or SSE
--server-url URL	Override the server (or env PREHACK_SERVER_URL; default https://prehack-server.fly.dev)
--ttl SECONDS	Maximum time to wait for approval
--dry-run	Show, do not write

Free mode vs Plug & Play

After connecting, you can use PreHack without a subscription. Be clear about what you get:

Free mode is not the PreHack pentester agent. It runs a set of deterministic platform tools (pattern matching and fixed checks): no AI, no proof of exploit, no fixes, no dynamic scan of your running app. Coverage and accuracy are lower than the agent and false positives are expected. Findings include file and line. Unlimited. In free mode nothing is ever written to your GitHub.

Plug & Play (US$29/mo) adds AI-powered detection (more vulnerability classes, fewer false positives), proof of exploit for every finding, fixes delivered as pull requests, and dynamic scans of your running app.

	Free mode	Plug & Play
Engine	Deterministic checks (no AI)	AI pentester agent
Proof of exploit	No	Every finding
Fixes as pull requests	No	Yes
Dynamic scan of the running app	No	Yes
Writes to your GitHub	Never	Only pull requests you enable: the report, or the fix (your choice of level)
Limit	Unlimited	See prehack.ai

The server tells your editor which mode you are in: tool results carry modo: "free" and a notice block while in free mode.

What is PreHack

PreHack is an autonomous, multi-agent AI pentester offered as a service. In short:

1. Connect GitHub. That connection is your consent to test your code.
2. It tests like an attacker, including dynamic scans of the running app.
3. It proves every finding with a real proof-of-concept attack in an isolated sandbox. No proof, no report.
4. It sends the fix as a Pull Request.

Steps 2 to 4 describe the pentester agent, which is the Plug & Play plan (US$29/month) with managed AI credit: PreHack runs the AI, you do not need an API key. Without a subscription you get free mode, which is deterministic checks only. See prehack.ai for details, current pricing and FAQ.

Security and trust

The wall. This public repository holds only the installer. By design it contains:

• no scanning engine and no agent orchestration;
• no security prompts and no attack or exploitation skills;
• no secrets, keys or tokens (the token is issued to you, at run time, and stored only in your editor config).

You can verify this: the whole connector is scripts/prehack_connect.py (stdlib only) and SKILL.md. Read them before you run them.

What the script does on your machine: talks to the PreHack server to run the device-code flow, and edits your editor's MCP config file (with backup). Nothing else.

Authorized use only. Test only code and targets that you own or have written permission to test. PreHack must not be used against third parties' systems. You are responsible for having the authorization and complying with applicable law.

Found a security problem in this connector? Email security@prehack.ai (see SECURITY.md) rather than opening a public issue with sensitive details. For anything else, reach us at hello@prehack.ai or via prehack.ai.

Testing

python3 -m py_compile scripts/prehack_connect.py
python3 scripts/prehack_connect.py --help
python3 -m unittest discover -s tests -v   # uses a local fake server, no network

The tests pass on Python 3.9 through 3.14. To check the oldest supported version with uv: uv run --python 3.9 --no-project python -m unittest discover -s tests.

Links

• Website, pricing and FAQ: prehack.ai
• Sign-in standard: RFC 8628, OAuth 2.0 Device Authorization Grant
• MCP (Model Context Protocol): modelcontextprotocol.io

License

Apache License 2.0. This license covers only this connector (the installer in this repository), not the PreHack service — the scanning engine and everything that powers it are proprietary and run on PreHack's servers.