MCP Integrations

Safeguard ships its own remote MCP (Model Context Protocol) server, so you can connect your favorite AI assistant directly to your software supply chain security data and ask questions in plain language. Once connected, your assistant can query your vulnerabilities, SBOMs, packages, licenses, findings, and compliance data, generate AI remediation plans, compare SBOMs, manage SCM integrations, and enforce policy gates. The server registers hundreds of tools and also supports MCP resources and prompts. Tool availability is governed per tenant — a curated set is enabled by default and administrators enable more per workspace (see Tool availability & gating). Agents can even onboard and procure Safeguard itself — create the account through the unauthenticated https://mcp.safeguard.sh/mcp/signup endpoint (signup and email verification only), sign in with OAuth on the main endpoint, browse regional pricing, compare tiers, start a Stripe checkout, verify activation, and read the user's notifications (see Payments & procurement and the Agents & LLMs guide).

Choose your assistant

Pick your AI assistant below for step-by-step connection instructions. The guides span both US and European model companies. Every page uses an endpoint on the mcp.safeguard.sh host.

Authentication

You need a Safeguard account first. Safeguard hosts the MCP server at https://mcp.safeguard.sh/mcp (Claude's connector URL is https://mcp.safeguard.sh/mcp/anthropic), so there is nothing to install. There are two ways to authenticate:

• OAuth 2.0 (recommended). Add the URL to your assistant. It sends you to Safeguard's sign-in page: sign in at https://app.safeguard.sh, check the app name and the host the page says the authorization code will be sent to, then click Authorize access. No keys to copy. Claude, ChatGPT, Cursor, VS Code and Gemini CLI all sign in this way.
• MCP key. For a client that can only send a fixed header. In Safeguard, open Settings → Developer, create a key of the type MCP key, and send it as Authorization: Bearer sg_sk_.... The key is shown once. If you don't see the Developer tab or the MCP key type, ask your Safeguard administrator: API-key access to MCP is turned on per workspace.

Either way, your assistant acts as you: it can see and do only what your Safeguard role allows, within the tools your workspace has turned on. Two things to know about MCP keys:

• An MCP key acts as the member who created it, across the whole workspace. Give it only to a client you trust, and keep it out of source control.
• Revoking takes effect within a minute. Revoke a key under Settings → Developer. The server checks each key at least once a minute, so a revoked key, a key whose owner's account is closed, or a key in a workspace whose API-key access is turned off stops working within a minute.

The server registers hundreds of tools for vulnerability scanning, SBOM analysis, AI-powered remediation, SCM integration, and policy enforcement. Which of them your assistant can see and call is controlled per tenant — see Tool availability & gating.

Example prompts

Once connected, plain-language prompts work the same regardless of which assistant you picked above — this is the same starter list used across every per-assistant guide:

List my security projects
Find critical vulnerabilities
Show me all SBOMs
Generate a remediation plan for my npm project
Which packages violate my license policy?

Enforcing policy on MCP tool calls

Connecting an assistant to Safeguard's MCP server and gating which of Safeguard's own tools it can see (the previous section) is one layer of control, and it only governs Safeguard's own bundled tool surface. Guard is a separate product that adds a second, complementary layer: live enforcement on MCP tool calls in general — including MCP servers and agents you build or operate yourself, not just Safeguard's — evaluated against your organization's live Safeguard policy in real time rather than a fixed allow-list.

Tool gating and Guard are independent mechanisms. Disabling one of Safeguard's own tools under Feature Flags does not create or change a Guard policy, and a Guard policy does not touch Safeguard's tool feature flags. Use Feature Flags to control access to Safeguard's own MCP tools; use Guard when you need the same kind of discovery-time/call-time control over MCP servers and agents you build or operate.

There are two ways to add Guard enforcement to an MCP server or agent:

• Guard SDK (not published yet): a library for enforcing your policy inside your own MCP server or agent process. There is no Guard SDK package on PyPI, npm or the Go module proxy, so there is nothing to install today, and any package that claims to be it did not come from Safeguard.

• Guard Proxy — run Guard as a standalone reverse-proxy gateway in front of one or more MCP servers you don't control the code of (or that are written in a language without a Guard SDK yet). Point your agent's MCP client at a proxy URL instead of the upstream server — POST https://<your-guard-host>/p/{slug}/message — and Guard evaluates every JSON-RPC call before forwarding allowed ones on.

Both approaches enforce the same underlying policy, authored in your Safeguard workspace rather than as files you check into source control. From there:

• Guard Policies covers how rules are authored: each one matches on method, tool name (glob wildcards), and/or agent ID, then resolves to allow, deny, or monitor, evaluated in priority order with first-match-wins. Rules can also carry input filters (block a pattern, cap request size, catch jailbreak phrasing) and output filters (mask PII in responses, redact fields, cap row counts) that run against tool traffic.
• Guard for servers & agents covers registering the MCP servers and agents in your environment as first-class records, so policy can target a specific server or agent, tool discovery can run against a registered server, and an individual agent can be blocked, unblocked, or quarantined directly.
• Guard Audit & Alerts covers the append-only trail Guard writes for every request it evaluates — allow, deny, and monitor alike — plus alert rules that watch signals like jailbreak score, anomaly score, and blocked-request ratio.
• Guard Webhooks covers delivering those alerts to your own endpoint as a signed HTTP webhook, separate from the platform-wide Webhooks & Events system.

FAQ & Troubleshooting

My assistant connects, but the tools menu is empty. First confirm the connection itself succeeded: you clicked Authorize access on Safeguard's sign-in page (OAuth), or your client sends a valid MCP key in the Authorization header. If the connection is fine, remember tool visibility is gated per tenant too: a fresh workspace only exposes a small curated default set until an administrator enables more. See Tool availability & gating.

I only see a handful of tools, not the full catalog. That's expected, not a bug. A new workspace starts with six read-only tools for browsing products, projects, vulnerabilities and findings; every other tool stays off until an administrator turns it on per tool under Feature Flags. See Tool availability & gating.

Should I use OAuth or an MCP key? Use OAuth wherever your client supports it, which includes Claude (web and Desktop), ChatGPT, Cursor, VS Code and Gemini CLI: there's no key to copy, store or rotate. Use an MCP key only for a client that can only send a fixed Authorization header.

My assistant authenticated fine yesterday, but every call fails today. If you connected with OAuth, sign in again from your client. If you use an MCP key, check whether it was revoked, whether the member who created it still has an active account, and whether your workspace still has API-key access to MCP: a key stops working within a minute of any of those changing. Create a new key under Settings → Developer if you need one.

Do I need SAFEGUARD_API_KEY, npx or an npm package? No. The server is hosted and never reads environment variables from your machine, and Safeguard publishes no package for it. Instructions that set SAFEGUARD_API_KEY or start a local bridge with npx are out of date: use the server URL with OAuth, or an MCP key in the Authorization header.

Do I need Guard as well as connecting my assistant, or is one enough? They solve different problems. Connecting your assistant plus tool gating controls what Safeguard's own MCP server exposes to that one assistant. Guard is a separate product for enforcing live policy on MCP traffic in general, including servers and agents that have nothing to do with Safeguard's own tool surface. Most teams only need Guard once they're building or operating MCP servers/agents of their own — connecting an assistant to Safeguard alone doesn't require it.

Related

• Tool availability & gating — the per-tenant feature flags that decide which of Safeguard's own tools an assistant can see and call.
• Guard SDK — embed live policy enforcement directly inside an MCP server or agent process you control.
• Guard Proxy — enforce policy in front of MCP servers you don't control the code of.
• Guard Policies — how allow/deny/monitor rules and input/output filters are authored.
• Servers & Agents — register the MCP servers and agents Guard policy can target.
• Audit Trail & Alerts — every decision Guard makes, plus alert rules for jailbreak, anomaly, and blocked-ratio spikes.
• Guard Webhooks — deliver Guard alerts to your own endpoint over a signed webhook.
• RBAC & Teams — the roles and permissions that decide what a connected assistant can see and do.
• SDKs: Safeguard publishes no language SDKs yet; the page lists what to use instead.