If your assistant has no dedicated guide, add a remote MCP server of type Streamable HTTP with the URL https://mcp.safeguard.sh/mcp. If the client supports OAuth for MCP servers, sign in when it asks. If it can only send a fixed header, use an MCP key. Safeguard has no SSE endpoint.
The generic recipe
- Create or sign in to a Safeguard account at app.safeguard.sh.
- Open your client's MCP configuration.
- Add a remote server, transport Streamable HTTP, URL
https://mcp.safeguard.sh/mcp. - If the client supports OAuth, it opens Safeguard's sign-in page. Check the app name and the host the code will be sent to, then click Authorize access.
- If the client can only send a fixed header, create an MCP key under Settings, then Developer, and send it as Authorization with the value Bearer followed by the key.
- Reload the client. Your enabled Safeguard tools appear.
Safeguard's documentation uses this same recipe for several assistants, including Perplexity, Mistral's Le Chat, Cohere, Aleph Alpha and Meta AI. The clients differ in menus, not in what you enter.
Addresses by client
| Client | Address |
|---|---|
| Claude | https://mcp.safeguard.sh/mcp/anthropic |
| Gemini CLI | https://mcp.safeguard.sh/mcp/gemini (field httpUrl) |
| Grok | https://mcp.safeguard.sh/mcp/grok |
| Cursor, ChatGPT, VS Code, others | https://mcp.safeguard.sh/mcp |
Cursor once used a Cursor specific path. It still works but the plain address is current.
Details that cause trouble
SSE versus Streamable HTTP. Some clients label these differently. Choose the option that means a single HTTP endpoint with streamed responses. If the only choice is SSE, the client is not compatible with the hosted endpoint.
Field names. Cursor uses mcpServers with url. VS Code uses servers with type http. Gemini CLI uses httpUrl. Copying between them without changing the names is the usual error.
Discovery. The server publishes OAuth metadata at the standard well known paths, which is how OAuth capable clients find the sign in. You should not need to paste any client ID, because clients register themselves automatically.
What you get after connecting
Tool availability is per tenant. A new workspace starts with six read-only tools, and administrators turn on more. See why only six tools show up.
What to avoid
Do not use npx, mcp-remote or any local bridge. Safeguard publishes none. Do not set SAFEGUARD_API_KEY. And do not put a key in a file you commit. If a client forces that, use a client that supports OAuth.
Testing the connection outside your assistant
The server exposes a health address and an OpenAPI description. You can open https://mcp.safeguard.sh/health in a browser to confirm the service is up before debugging your client.
Click through it at your own pace, or open it full screen.
Quick check
My client only offers SSE. Can I still connect?
Not to the hosted endpoint. Safeguard has no SSE endpoint, so you need a client or setting that supports Streamable HTTP.
Do I need a different account per assistant?
No. One Safeguard account works. Each client authorizes separately.
Sources
- Safeguard docs: MCP integrations overview
- Safeguard docs: Gemini guide
- Safeguard docs: Grok guide
- Safeguard docs: Perplexity guide
- Safeguard MCP server descriptor
Last reviewed: 19 October 2026.
