GitGuardian MCP Server

The GitGuardian MCP Server brings GitGuardian's secrets detection straight into your AI assistant. Secrets detection means finding credentials such as API keys, database passwords or cloud tokens in code before they reach a public repository or the wrong hands. MCP, the Model Context Protocol, is an open standard that assistants such as Claude Code, Cursor or Claude Desktop use to call external tools. The server comes from GitGuardian itself, is published on GitHub as an open source project under the MIT licence and is built on the GitGuardian API. According to the README, it relies on more than 500 detectors for different kinds of secrets. The official documentation still labels the server as beta, so features may change.

Why this matters with AI assistants in particular

AI assistants now write large parts of configuration files, scripts and sample code. It is easy for a real key to slip into a file that gets committed later. The value of the GitGuardian MCP Server is that the assistant takes on this check itself: you ask it to search the codebase for leaked credentials, and it calls GitGuardian's detection engine for that instead of guessing from suspicious patterns. Unlike simple regex checks in the editor, the server uses the same detection that GitGuardian runs in its platform and in ggshield. And it knows the incidents already recorded in your workspace.

What can the GitGuardian MCP Server do?

According to the README, the server covers four areas. First, scanning files for leaked credentials. Second, incident management: the assistant lists, filters, counts and opens incidents, assigns them to team members, adds tags and notes and marks them as resolved. This applies to internal incidents as well as findings from Public Monitoring, the surveillance of public sources. Third, honeytokens: the assistant generates new decoy credentials and lists existing ones. Fourth, code-fix automation: for repositories your workspace monitors, the server can open pull requests that remove a detected secret from the code. The repository also contains tools for fetching remediation workflows, detectors, sources, teams and members, as well as a tool for revoking secrets. Which of them your assistant actually sees depends on the OAuth scopes of your access token.

Hosted, self-run or local

The README recommends the hosted server. For accounts in the US region the address is https://mcp.gitguardian.com/mcp, for the EU region https://mcp.eu1.gitguardian.com/mcp. The MCP client handles the OAuth sign-in on its own; you install nothing locally and manage no token. If you run GitGuardian yourself, work in an air-gapped environment or want the server on your own infrastructure, use the Docker image ghcr.io/gitguardian/mcp-server behind a reverse proxy. According to the README, the server is soon to ship directly in the Helm chart of the self-hosted installation. For CI/CD pipelines, scripts and older clients without HTTP support, there is a local stdio mode via uvx with a personal access token.

Security and permissions

GitGuardian states plainly in the README that agents act on your behalf and under your responsibility. According to the vendor, the server therefore leans towards read permissions by default. What is actually unlocked is determined by the token's scopes. Tokens issued through OAuth are valid for 30 days by default. With the hosted server, keep in mind that requests pass through mcp.gitguardian.com in addition to the GitGuardian API. The older browser sign-in in stdio mode is marked as deprecated and is set to be removed.

Cost

The server itself is free and open source. You need a GitGuardian account. According to the pricing page, there is a permanently free Starter plan for smaller teams with internal secrets monitoring and real-time scanning. Public Monitoring, honeytokens, SSO and other features are reserved for paid Business, Growth or Enterprise plans. You will find current terms on GitGuardian's official pricing page.

Who is the GitGuardian MCP Server for?

It pays off for teams that already use GitGuardian and want to work through security incidents faster without jumping between dashboard, editor and ticketing system. It is just as interesting for developers who let AI assistants generate a lot of code and want a reliable credentials check before every commit. Start with a token that has few scopes and only widen its permissions once you trust the workflow.

A GitGuardian account (the free Starter plan is enough to get going) and an MCP client with HTTP transport, for example Claude Code, Cursor or Claude Desktop. For local mode, additionally uv/uvx and a personal access token.

With the GitGuardian MCP Server, an AI assistant scans code for leaked credentials and works through security incidents from GitGuardian. This guide leads from choosing the region to the first scan.

Step 1: Prerequisites

You need a GitGuardian account and an MCP client with HTTP transport, such as Claude Code, Cursor or Claude Desktop. According to the pricing page, the free Starter plan is enough to get going. Check first whether your workspace lives in the US or the EU region. The dashboard address tells you: dashboard.gitguardian.com means the US, dashboard.eu1.gitguardian.com means the EU.

Step 2: Connect it to Claude Code

For US accounts the command is claude mcp add --transport http gitguardian https://mcp.gitguardian.com/mcp. For EU accounts, register https://mcp.eu1.gitguardian.com/mcp instead. In Cursor, Windsurf or Zed, add the same address with type http to the relevant mcp.json. In Claude.ai in the browser, add the server as a custom connector under Settings and Connectors. Then check with /mcp that the server is connected.

claude mcp add --transport http gitguardian https://mcp.gitguardian.com/mcp

Step 3: Sign in

On first use, the client starts the OAuth sign-in in the browser. GitGuardian then issues an access token that is valid for 30 days by default. Which tools the assistant sees afterwards depends on the scopes of that token. If a tool you expect is missing, it is usually down to missing permissions, not the installation.

First scan

Ask the assistant something like: Scan this codebase for leaked credentials. It sends the relevant files to the GitGuardian API for checking and reports findings with the file name and the kind of secret. Treat any key it finds as compromised. Do not just remove it from the code, rotate it with the service it belongs to.

Work through incidents

With questions like: Are there new incidents assigned to me? the assistant fetches the open incidents from your workspace. It can explain them, suggest remediation steps, assign them to team members and mark them as resolved. For monitored repositories, the server can also open a pull request that removes the secret from the code. Review such pull requests like any other code change.

Local or in CI without OAuth

For pipelines and older clients, the server runs locally via uvx using the package from github.com/GitGuardian/ggmcp and the command gg-mcp-server. Set ENABLE_LOCAL_OAUTH=false and put a personal access token from the dashboard under API and Personal Access Tokens into GITGUARDIAN_PERSONAL_ACCESS_TOKEN. In Claude Desktop on macOS, the entry needs the absolute path to uvx.

OAuth via the hosted server (tokens valid for 30 days by default), a personal access token in GITGUARDIAN_PERSONAL_ACCESS_TOKEN (local/CI) or a bearer token in the Authorization header (self-run HTTP server).

Depending on token scopes: scan files for secrets, read, assign, tag and resolve internal and public incidents, generate and list honeytokens, open code-fix pull requests and revoke secrets.

File contents to be scanned are sent to the GitGuardian API; with the hosted server, requests additionally pass through mcp.gitguardian.com. Incident data that is read enters the AI client context and therefore reaches the model provider.

Incident data can point to real secrets, honeytoken values are returned unmasked, with broad scopes incidents can be closed or secrets revoked, and personal access tokens sit in configuration files in local mode.