If you already work inside an AI assistant, you should not need to leave it to ask about your dependencies. Safeguard runs a hosted MCP server at mcp.safeguard.sh, so Claude, ChatGPT, Cursor or an agent you build can query your supply chain data in plain language.
Tool calls are unlimited on every plan. What an assistant can actually do is decided by your role and by which tools your workspace has turned on.
What it is
The Model Context Protocol, MCP, is a standard way for an assistant to call tools. Safeguard hosts the server, so there is nothing to install on your machine. Once connected, your assistant can query vulnerabilities, SBOMs, packages, licenses, findings and compliance data, generate remediation plans, compare SBOMs, and check policy gates.
You also keep your model choice. You connect Safeguard to Claude or ChatGPT through MCP and bring your own model. This is a connection through MCP, not a pass-through of any assistant subscription.
Before you start
You need a Safeguard account. Create one at app.safeguard.sh or sign in if you have one.
The docs are firm on one point: Safeguard publishes no npm package and no local bridge for its MCP server. Do not run anything that claims to be the Safeguard MCP server, and do not set an API key environment variable for it. The server never reads variables from your machine. Treat any such instruction as out of date or hostile.
How to use it
Claude
The connector URL for Claude is https://mcp.safeguard.sh/mcp/anthropic.
Shortest route: Safeguard is listed in Claude's connector directory. Select Safeguard and connect, sign in to Safeguard, check the app name and the host the authorization code will be sent to, then click Authorize access.
Manual route:
- In Claude, open Customize, then Connectors, click Add, then Add custom connector.
- Name it Safeguard Security and enter the URL above. Click Continue.
- Claude detects OAuth. Choose Sign in now or Sign in when needed, and choose to register the OAuth client automatically. Click Add.
- When Claude sends you to Safeguard, sign in, check the host shown, and click Authorize access.
On Team or Enterprise plans, an Owner adds the connector under organization settings first, and members connect afterwards. Claude Desktop uses the connectors from your account, so do not edit claude_desktop_config.json for Safeguard. That file is for local servers, and this one is hosted.
Cursor
Cursor connects to https://mcp.safeguard.sh/mcp. Add this to ~/.cursor/mcp.json for all projects, or .cursor/mcp.json inside one project:
{
"mcpServers": {
"safeguard": {
"url": "https://mcp.safeguard.sh/mcp"
}
}
}
Cursor supports OAuth for MCP servers. It opens Safeguard's sign-in page, you sign in, check the host, and click Authorize access.
ChatGPT
The recommended path is a custom MCP server at https://mcp.safeguard.sh/mcp with OAuth.
- In ChatGPT, open Plugins, select the plus button, then Add custom MCP server.
- Enter a name, for example Safeguard Security, and a description.
- Under Connection, enter the server URL above.
- Choose OAuth. ChatGPT registers itself, so there is no client ID or key to copy. Review the risk warning, select I understand and want to continue, then select Create as a plugin.
- When ChatGPT sends you to Safeguard, sign in, check the host shown, and click Authorize access.
Account and workspace policies apply to custom MCP servers in ChatGPT. If you do not see the option, ask your ChatGPT workspace administrator.
Your own agent
Point any MCP client at https://mcp.safeguard.sh/mcp. If it can run OAuth, use that. If it can only send a fixed header, use an MCP key: in Safeguard open Settings, then Developer, and create a key of type MCP key. It looks like this when sent:
Authorization: Bearer sg_sk_...
The key is shown once. An MCP key acts as the member who created it, across the whole workspace, so keep it out of source control. Revoking it under Settings, then Developer, takes effect within a minute. If you do not see the Developer tab, ask your admin, because API key access to MCP is turned on per workspace.
What your assistant will see
Do not be alarmed by a short tool list. A new workspace starts with six read-only tools: list products, list projects, get a project, list vulnerabilities, get a vulnerability and list findings. Nothing in that set can create, change or delete anything.
Every other tool is off until an administrator turns it on, one tool at a time, under Feature Flags. A tool that is off does not appear in the tool list at all, and a call to it returns a normal error result saying it is not enabled for the tenant.
That design is useful for review. You can start read-only, watch what your assistant does with it, and add scanning or integration tools only when you want them.
Troubleshooting
- Connected, but the tools menu is empty. Confirm you clicked Authorize access. Then check which tools are enabled.
- Worked yesterday, fails today. Sign in again from your client. For an MCP key, check that it was not revoked and that its owner still has an active account.
- A tool is missing. Copy its name from the error message and search for it under Feature Flags.
Try it yourself
Connect one assistant. Ask it to list your projects, then to find critical vulnerabilities. Ask for a tool you know is disabled and read the error message. Then ask your admin to enable one more tool and repeat the request.
Click through it at your own pace, or open it full screen.
Quick check
Do you need to install an npm package to use the Safeguard MCP server?
No. The server is hosted, and Safeguard publishes no npm package or local bridge for it.
What can a new workspace's assistant do by default?
Six read-only tools: list products, list projects, get a project, list vulnerabilities, get a vulnerability and list findings.
When should you use an MCP key instead of OAuth?
Only when the client can send a fixed header and cannot run OAuth. OAuth needs no key to copy or rotate.
