UTC

MCP Server

A hosted remote MCP server with seven read-only lookup tools. No tool needs an account, and none changes anything.

Endpoint

URL
https://api.shieldpi.io/mcp/
Transport
Streamable HTTP, stateless
Authentication
None
Tools
Seven, all annotated read-only

The ShieldMap tools answer from ShieldMap, the dataset ShieldPi publishes of security advisories for MCP servers and SDKs, coding agents, agent frameworks and other AI-stack packages, and of the AI security provisions in regulations and standards. The other three describe ShieldPi's scanner. Every tool carries readOnlyHint: true and destructiveHint: false.

Add it to Claude

  1. Open Settings → Connectors.
  2. Choose Add custom connector.
  3. Name it ShieldPi and paste https://api.shieldpi.io/mcp/ as the URL. Leave the authentication settings empty: the server has no sign-in.
  4. Turn the connector on in a chat, then ask one of the questions below.

If your organization manages connectors, an owner may need to add it first. In Claude Code:

claude mcp add --transport http shieldpi https://api.shieldpi.io/mcp/

Any other MCP client that speaks Streamable HTTP can use the same URL.

Example prompts

  • “My agent uses mcp-remote 0.1.15 from npm. Does that version have known vulnerabilities, and what should I upgrade to?”

    Calls shieldmap_check_package, which returns CVE-2025-6514, a critical OS command injection, and the fixed release, 0.1.16.

  • “Look up CVE-2025-6514. Which packages does it affect, and is it exploited in the wild?”

    Calls shieldmap_advisory, which returns the advisory, the affected versions of mcp-remote, the fix, and whether CISA lists it as exploited.

  • “What does Article 15 of the EU AI Act say about cybersecurity?”

    Calls shieldmap_requirements, which returns Article 15 quoted paragraph by paragraph, including 15(5) on resilience against attempts to exploit system vulnerabilities, with links to the source.

  • “Which LangChain packages are on ShieldMap, and which have the most advisories?”

    Calls shieldmap_search, which returns the matching npm and PyPI packages, the most advisories first.

  • “How does ShieldPi decide that a scan finding is real?”

    Calls verification_methodology, which returns the judging rules for model scans and the limits of the public MCP probe.

ShieldMap tools

shieldmap_check_package

Check a package on ShieldMap

The published vulnerability and malicious-package advisories for one package. With a version, only the advisories that version is affected by, and the release to upgrade to. For advisories no release fixes, mitigations lists the advisory's own workaround text, with its source, and ShieldPi's general guidance for the kind of weakness, labelled as not a vendor fix. A package ShieldMap does not track returns on_shieldmap: false, which means ShieldMap has no data on it, not that it is safe.

ParameterRequiredAccepts
ecosystemYesnpm, pypi, go, maven, rust, nuget, rubygems, composer or swift
nameYesThe package name as the registry spells it, e.g. mcp-remote (up to 214 characters)
versionNoThe installed version, e.g. 0.1.15 (up to 64 characters)

shieldmap_search

Search ShieldMap packages

Up to 20 packages whose name contains the query, those with the most advisories first, each with its advisory count and whether an advisory marks it malicious or exploited in the wild.

ParameterRequiredAccepts
queryYesA word or part of a package name, 2 to 100 characters

shieldmap_advisory

Look up a ShieldMap advisory

One advisory: its summary and severity, the affected packages and version ranges, the fixed versions, whether CISA lists it as exploited in the wild, and the link to the original advisory. Where no release fixes it, the same mitigations as shieldmap_check_package. An ID ShieldMap does not have returns advisory: null.

ParameterRequiredAccepts
advisory_idYesA GHSA ID such as GHSA-6xpm-ggf7-wc3p, or a CVE ID such as CVE-2025-6514

shieldmap_requirements

Quote AI security requirements

Quoted text from the EU AI Act, NIST AI 600-1, the NIST AI RMF, the CoSAI Risk Map and the OWASP Top 10 for LLM Applications 2026 crosswalk on an id or a topic, with links to ShieldMap and the source and each result's licence. EU AI Act text comes from its current consolidated version; a question about its dates also returns the application dates from Article 113. Quotations to navigate by, not legal advice.

ParameterRequiredAccepts
queryYesAn id such as Article 15(5), Annex III, LLM05, MEASURE 2.7 or riskDataPoisoning, or a few words such as data poisoning

Scanner tools

scan_modes

ShieldPi scan modes

The scan modes this deployment supports and the kind of target each one tests.

No parameters.

platform_stats

ShieldPi platform summary

The attack corpus version, the number of scan modes, how scan results are judged, what the standards mappings do and do not establish, and this server's MCP protocol versions and authorization model.

No parameters.

verification_methodology

ShieldPi verification methodology

How ShieldPi verifies findings and the limits of each method, with a link to the published methodology.

No parameters.

This server does not run scans. To grade a public MCP server, use the MCP scan page.

Limits and errors

  • Request bodies over 256 KiB are refused with HTTP 413.
  • The server takes no credentials. A URL carrying token, access_token or api_key is refused with HTTP 400.
  • Browser requests from other websites are refused with HTTP 403.
  • An argument that does not fit its parameter returns an error result that names the parameter.
  • ShieldMap data is read from the files ShieldPi publishes at www.shieldpi.io/map/data/ and kept for up to an hour. If they cannot be read and nothing is kept, the tool returns an error result that asks you to try again in a few minutes.
  • ShieldMap covers the packages and advisories it lists. A package or advisory it does not list has no data there, which does not mean it has no vulnerabilities.

Data handling

The server receives the tool name and arguments your client sends, and nothing else from your conversation. It answers from published data and does not store the arguments or results. The ShieldMap tools download whole files, so your query is not sent on to another service.

As for every API request, the server log records the client IP address, the request method and path, and the response status; credentials sent in a URL are masked. Request counts by path and status are kept for monitoring. See our privacy policy.

Your account server

Everything above describes the public server. A second one, at https://api.shieldpi.io/mcp/account, acts for your ShieldPi account: you sign in to connect it, and approve what it may do.

  1. Open Settings → Connectors and choose Add custom connector.
  2. Name it ShieldPi account and paste https://api.shieldpi.io/mcp/account as the URL.
  3. Under OAuth client, keep Use Claude's published identity, the recommended choice: ShieldPi reads Claude's details from the client metadata document Anthropic publishes, and the consent page names the host it read them from. Register automatically works too; Claude then registers a client with ShieldPi first. Leave the client ID and secret empty either way.
  4. Click Connect, sign in to ShieldPi, and check the consent page: who is asking, where you return afterwards and what it may do. Approving asks for your password.
ScopeLets the app
shieldpi:readRead your alerts, incidents, scans, reports, monitoring and guardrails
shieldpi:writeStart scans, acknowledge and resolve alerts, and place or release Session Holds
  • An app that asks for both scopes can be allowed read only, and one that names no scope gets shieldpi:read.
  • A connection lasts 90 days from approval, however often it refreshes. Disconnect it at any time under Settings → Connected apps. A password reset, or adding or removing two-factor authentication, ends it too.
  • Each connection may make 120 requests a minute, and each account 20 write calls a minute in all.

Other MCP clients

The authorization server's metadata is at https://api.shieldpi.io/.well-known/oauth-authorization-server. Clients use the authorization code flow with PKCE (S256), and refresh tokens change on every use. A client can name itself with the https URL of a client metadata document (draft-ietf-oauth-client-id-metadata-document). ShieldPi fetches it only from public addresses, follows no redirects, and keeps it for five minutes to a day, as its cache headers say. Such a client must be public, with token_endpoint_auth_method none, and must return to a redirect URI its document lists. A client can also register at /oauth/register (RFC 7591).

Support

Questions and problem reports: [email protected].