Connect your AI coding tool to MAGDOX and it can send the code it just wrote for scanning and get back what is vulnerable, where and how to fix it, before anything is committed, once an admin turns the scan tools on. It can also work through your security findings with MAGDOX's guidance in front of it, check a package name before installing it and start a scan of a repository connected through the GitHub App. It never ignores or changes a finding.
Connect
- In your AI tool, add a remote MCP server (often called a custom connector) with the address https://api.magdox.io/mcp. In Claude Code, run claude mcp add --transport http magdox https://api.magdox.io/mcp.
- The tool opens a MAGDOX page. Sign in, check which app is asking, where it will return and which access it wants, then approve with a fresh code from your authenticator app.
- The connection appears under Settings, Integrations, AI coding tools, where you can disconnect it at any time.
What the assistant can do
| Tool | What it does |
|---|---|
| scan_files | Scans up to 20 files (200 KB in all) the assistant sends and answers each finding with its file, line, what is wrong and how to fix it. The results are not added to your repository's findings. Off until an admin turns it on |
| scan_dependencies | Checks dependency manifests and lockfiles the assistant sends for known vulnerable and malicious packages, with the fixed version to upgrade to. Off until an admin turns it on |
| scan_result | Reads the result of a scan the assistant started; results are kept for an hour |
| list_projects | Lists your repositories |
| list_findings, get_finding | Reads findings and their details |
| explain_finding | Returns MAGDOX's written guidance: what the rule found, why it matters and how to fix it |
| start_scan, scan_status | Starts a scan of a repository connected through the GitHub App and follows it |
| check_package | Checks a package name before installation against known malicious and made-up names, and against names one character away from widely used packages |
Who stays in control
- Super Admins and admins choose which tools are allowed in the organisation under Settings, Integrations, AI coding tools. The two scan tools start off. Each scan an AI coding tool starts counts toward the organisation's limit of 200 hosted scans in any 24 hours, and one person can start up to 30 scans with scan_files or scan_dependencies in an hour. No tool can ignore or change a finding.
- Access is read findings, or read findings and start scans, for one organisation and the person's role; only a person with the SuperUser role can give start scans, which start_scan, scan_files, scan_dependencies and scan_result need. It never reaches settings, members or billing.
- Only known AI tool addresses (claude.ai, claude.com, chatgpt.com, vscode.dev and Cursor) and tools on the person's own computer can receive an approval.
- Every call is recorded in the audit log. Access tokens last an hour and refresh tokens rotate on every use. A connection ends when the person leaves the organisation, changes or resets their password, changes their email or removes an authenticator, after 90 days without use, and a year after it was approved; connect again from the tool. One person can hold up to 20 connections and add up to 10 in 24 hours.
- What the assistant reads goes to its own provider under your agreement with that provider.
An AI coding tool is separate from your organisation's own AI model, which an admin connects under Settings, Integrations, AI model to write recommendations, fixes and regression tests on findings.