To use Safeguard from Claude, add Safeguard from Claude's connector directory and sign in once with OAuth. To use it from ChatGPT, install Safeguard from the ChatGPT plugin directory. After that, you can ask about projects, vulnerabilities and compliance in the chat you already use, with the access your admin allows.
TL;DR
- Safeguard runs as a Model Context Protocol (MCP) server, so AI assistants can talk to it.
- In Claude, add it from the connector directory on web, desktop or mobile, and sign in with OAuth. There is no API key to paste.
- In ChatGPT, install it from the plugin directory, or import the schema address shown on the download page.
- Other assistants that support MCP can use the same endpoint.
- What the assistant can see and do depends on the access your admin grants.
Definition: MCP, the Model Context Protocol, is a shared standard that lets an AI assistant call outside tools and read outside data. An MCP server is the program on the other side that offers those tools. Think of it as a universal plug between a chat assistant and a service.
Why would I connect Safeguard to a chat assistant?
Most people already have a chat assistant open all day. Opening another dashboard to ask "do we have any critical CVEs?" is one more tab to find. The Safeguard demo on this topic puts it this way: your team already works in the assistant, so connect it, and people can ask about projects, vulnerabilities and compliance in the chat they use every day.
A CVE is a public ID for a known software flaw. Asking in plain English is faster than building a filter, especially for people who do not use security tools much, such as a product manager or a founder.
The download page shows the sort of question it has in mind: "Any critical CVEs?" The assistant turns that into a request to Safeguard, reads the answer, and explains it to you.
This is a convenience layer, not a different product. The same findings and the same permissions sit behind it.
How do I connect Claude?
The Safeguard download page lists this path:
- Open Claude on the web, on desktop or on mobile.
- Open the connector directory and add Safeguard.
- Sign in to Safeguard when asked. The page says you sign in once with OAuth and there is no API key to create or paste.
- Start a chat and ask a question about your projects.
OAuth is the sign in method that lets one app prove who you are to another without sharing your password. You approve, and the assistant gets a limited token.
If you prefer to add the connector by its address, the download page gives this MCP endpoint:
https://mcp.safeguard.sh/mcp
The demo linked at the end of this post shows Safeguard's side of the setup, so a workspace admin can see what to expect.
How do I connect ChatGPT?
The download page says to install Safeguard from the ChatGPT plugin directory. Then every scan, SBOM and advisory tool becomes available in the conversation. An SBOM is a software bill of materials, the ingredient list of a piece of software.
If you need to set it up by hand, the page also gives a schema address you can import:
https://mcp.safeguard.sh/openapi.json
A schema is a machine readable description of what a service can do. Importing it tells the assistant what questions it can send.
What about other assistants?
The download page says Safeguard is a single MCP server with a ready made configuration for each client, and that coding assistants and editors that support MCP can use it. The same endpoint above works for any MCP client. The page offers a "pick your client" option for the exact settings.
For readers who want to understand the standard itself, see our posts on what MCP means and on MCP security.
What can I ask?
Start with simple, read style questions, such as:
- "Which of my projects have critical vulnerabilities?"
- "Tell me about this CVE and whether a fix exists."
- "Which packages are in this project?"
- "What compliance controls are failing?"
The download page says an assistant can scan a repository, read an SBOM and check an advisory without leaving the conversation. What you can actually ask depends on which tools your admin has turned on for you.
Good prompts name the project and the severity. "Show critical vulnerabilities in the checkout service" gets a better answer than "what is wrong?"
How do I keep this safe?
An assistant that can call tools is more powerful than one that only chats. A few habits keep it safe.
| Habit | Why it helps |
|---|---|
| Use the least access that works | A person who only reads reports does not need tools that change things |
| Ask your admin what is allowed | The demo says access follows what your admin allows |
| Treat answers as leads | Check a surprising answer in the platform before you act |
| Be careful with pasted content | Text from outside sources, such as a web page, may try to steer an assistant. This is called prompt injection |
| Keep a record | Log who asked what, especially if the assistant can take actions |
Prompt injection means hidden instructions inside content that an assistant reads. If a connected assistant can change things, a trick like that is more serious. That is why teams ask for guardrails on tool use. Safeguard offers MCP Guardrail and agent identity as generally available capabilities, and you can read how they fit together in our post on MCP Guardrail and agent identity.
What does a chat session look like, with invented numbers?
The numbers below are invented to show a typical session.
A team lead opens Claude on Monday and asks, "How many critical vulnerabilities are open across our products?" The answer says 9 across 3 products (made up for this example). She asks, "Which of those have a fix available?" The assistant says 6 do.
She then asks for the top 3 by exposure. The assistant lists them with the package name and fixed version. She copies the list into a ticket for her team. Total time: about two minutes, with no new dashboard to learn.
Notice what she did not do. She did not ask the assistant to change anything. She used it to read and summarize, then handed real work to her team. That is a safe way to begin.
How does this fit with the rest of Safeguard?
Chat is one of several ways in. The download page also lists the desktop app, the editor extensions, the browser extensions and the CLI. See our guide to choosing a surface to pick the right one.
Safeguard offers generally available capabilities that sit behind the chat connector: SBOM and dependency scanning, reachability analysis, zero day discovery, policies and gates, autonomous remediation with fix pull requests (off by default, every category on Manual), MCP Guardrail, agent identity and compliance evidence automation. You do not need any of that to try the free scan at gold.safeguard.sh/scan, which works with no login.
The Safeguard CLI is another place where an assistant lives. It runs on your machine, and its assistant can also write code, asking before it changes files.
Try it yourself
Click through it at your own pace, or open it full screen.
Click through the seven steps and look for what the admin sets up on Safeguard's side before anyone asks a question.
Quick check
Do I need an API key to connect Claude?
No. The download page says you sign in once with OAuth, and there is no API key to create or paste.
Can ChatGPT use Safeguard too?
Yes. The download page says to install Safeguard from the ChatGPT plugin directory. It also lists a schema address you can import if you set it up by hand.
Can the assistant see everything in my account?
It sees what your admin allows. The demo describes the connection as working with the access your admin grants, so ask your admin which tools are turned on for you.
Is it safe to let an assistant call security tools?
It can be, if you limit access, check surprising answers and keep logs. Start with read style questions before you allow anything that changes data.
Sources
- Safeguard download page
- Use Safeguard from Claude demo
- Use Safeguard from ChatGPT demo
- Connect the Safeguard MCP server demo
- Safeguard Gold, Free Dependency Scan
Last reviewed: 19 October 2026.
