Socket MCP Server

A Model Context Protocol (MCP) server for Socket integration — lets AI assistants query dependency vulnerability scores and security metadata.

Socket MCP exposes Socket.dev’s package-scoring API through the Model Context Protocol, so any MCP-aware AI assistant (Claude, VS Code Copilot, Cursor, Windsurf) can score a package, audit a package.json, or flag risky dependencies as part of a conversation. It ships as both a hosted public server (https://mcp.socket.dev/, no setup) and a self-hostable npm package, so you can choose between zero-friction and full data isolation.

✨ Features

• 🔍 Dependency Security Scanning - Get comprehensive security scores for npm, PyPI, cargo, Maven, NuGet, RubyGems, Go Modules, and more (supported ecosystems)
• 🌐 Public Hosted Service - Use our public server at https://mcp.socket.dev/; sign in once via OAuth, no self-hosting
• 🚀 Multiple Deployment Options - Run locally via stdio, HTTP, or use our service
• 🤖 AI Assistant Integration - Works seamlessly with Claude, VS Code Copilot, Cursor, and other MCP clients
• 📊 Batch Processing - Check multiple dependencies in a single request
• 🔒 OAuth Sign-In - Public server authenticates through your MCP client’s OAuth flow; no API key to copy or manage

🛠️ This project is in early development and rapidly evolving.

Install

Option 1: Use the public Socket MCP server (recommended)

The easiest way to get started. The public server uses OAuth — your MCP client opens a browser to sign in to Socket on first connect; no API key to copy or manage. Click a button below to install in your favorite AI assistant.

Option 2: Self-host the Socket MCP server

Self-hosting keeps every request inside your own infrastructure. It needs a Socket API token and Node.js 24 or later.

Get a token first. Sign in at socket.dev, open the API tokens page, and create a token with the packages:list scope. That single scope covers depscore; the org-scoped tools need whatever scopes your organization requires for the endpoints they call. Full walkthrough: creating and managing API tokens.

Then pick a transport. The server speaks the same MCP protocol either way; the difference is who launches it.

Stdio is the default and the right answer for a single developer. Choose HTTP when more than one person, or something that is not a local process, needs to reach the server.

Enabling OAuth. Set all three OAuth variables to require an OAuth access token on every MCP request:

MCP_HTTP_MODE=true \
SOCKET_OAUTH_ISSUER=https://issuer.example.com \
SOCKET_OAUTH_INTROSPECTION_CLIENT_ID=your-client-id \
SOCKET_OAUTH_INTROSPECTION_CLIENT_SECRET=your-client-secret \
npx @socketsecurity/mcp@latest --http # socket-lint: allow npx

With OAuth enabled, every request to the MCP endpoint goes through RFC 7662 token introspection. A raw Socket API token is rejected with a 401, whatever its prefix; a sktsec_ token is no more privileged than any other string here. Callers send an OAuth access token, and the server uses that token for the Socket API calls it makes on their behalf. Run without the OAuth variables to accept raw Socket API tokens instead.

At startup the server discovers the issuer’s RFC 8414 metadata. An issuer that carries a path (https://auth.example.com/tenant1) is probed at the path-inserted well-known URL first (https://auth.example.com/.well-known/oauth-authorization-server/tenant1), and the metadata document’s own issuer must match SOCKET_OAUTH_ISSUER byte for byte. The issuer must be https on a public host; a loopback or private-network issuer is refused outright unless SOCKET_DEBUG=true is set for local-stack work. The same rule applies to the introspection_endpoint the metadata advertises.

Clients discover how to authenticate from GET /.well-known/oauth-protected-resource (RFC 9728), which the server publishes once OAuth is on:

{
  "resource": "https://mcp.example.com/",
  "authorization_servers": ["https://issuer.example.com"],
  "scopes_supported": ["packages:list"],
  "bearer_methods_supported": ["header"],
  "resource_name": "Socket MCP Server"
}

A rejected request answers 401 with a WWW-Authenticate header naming that metadata URL and the scopes this resource requires:

WWW-Authenticate: Bearer error="invalid_token", error_description="Invalid or expired token",
  resource_metadata="https://mcp.example.com/.well-known/oauth-protected-resource", scope="packages:list"

[!IMPORTANT] Audience validation. When an introspection response includes an aud claim, the server requires it to name this resource and rejects the token otherwise. That check is always on and has no opt-out. When the response includes no aud at all, the token is accepted by default, because an authorization server that never emits the claim would otherwise fail every request. Audience validation therefore protects you exactly as far as your authorization server populates aud; if it does not, no audience is being checked. Set SOCKET_OAUTH_REQUIRE_AUDIENCE=true to additionally require the claim’s presence, but only once you have confirmed your authorization server returns aud on introspection. Enabling it against a server that omits the claim rejects all traffic.

Add TRUST_PROXY=true only when the server is deployed behind a trusted reverse proxy or load balancer that normalizes the forwarded host and protocol headers.

Configure your MCP client to connect to the HTTP server:

{
  "mcpServers": {
    "socket-mcp": {
      "type": "http",
      "url": "http://localhost:3000"
    }
  }
}

Usage

Once installed, ask your AI assistant questions like:

• “Check the security score for express version 4.18.2”
• “Analyze the security of my package.json dependencies”
• “What are the vulnerability scores for react, lodash, and axios?”

Tools exposed

depscore

Query the Socket API for dependency scoring information. Returns supply chain, quality, maintenance, vulnerability, and license scores per package.

Supported ecosystems

Based on Socket’s language support. The ecosystem parameter maps to PURL types:

packagist is accepted as an alias for composer, and openvsx for the vscode PURL type.

Example request:

{
  "packages": [
    { "ecosystem": "npm", "depname": "express", "version": "4.18.2" },
    { "ecosystem": "pypi", "depname": "fastapi", "version": "0.100.0" }
  ]
}

Response, as a single block of text. Every score is an integer from 0 to 100, higher is better:

Dependency scores:
pkg:npm/[email protected]: license: 100, maintenance: 87, quality: 100, supplyChain: 97, vulnerability: 98
  Report: https://socket.dev/npm/package/express
pkg:pypi/[email protected]: license: 100, maintenance: 100, quality: 100, supplyChain: 100, vulnerability: 100
  Report: https://socket.dev/pypi/package/fastapi

Packages come back in the order the API returns them, not the order you asked for. A package Socket has no record of is left out of the list rather than reported as an error, so compare the response against your request when a name is missing.

organizations

List the Socket organizations the authenticated user belongs to. Takes no parameters. Use it to discover the org_slug value that the org-scoped tools (alerts, threat_feed) require.

This tool needs a Socket API token. See Authentication for organization-scoped tools below.

The response is the Socket API’s JSON, keyed by organization id. The slug is what the other tools want:

{
  "organizations": {
    "1234": {
      "id": "1234",
      "name": "Acme Robotics",
      "plan": "enterprise",
      "slug": "acme-robotics"
    }
  }
}

alerts

List the latest security alerts for one Socket organization: supply-chain, vulnerability, quality, license, and maintenance issues across the org’s monitored packages. Backed by GET /v0/orgs/{org_slug}/alerts. Results are paginated; pass the previous response’s endCursor as cursor to fetch the next page.

threat_feed

Look up items in a Socket organization’s threat feed: packages recently flagged as malware, typosquats, obfuscated code, and similar. Backed by GET /v0/orgs/{org_slug}/threat-feed. The response carries a nextPageCursor; pass it as cursor to page forward.

package_files

List the files published in a package: a tree of file paths, each with its size and blob hash, for any package on a supported ecosystem. Use it to inspect what a dependency ships before installing it. Pass a file’s hash to package_file_contents or package_file_grep.

Output is a header line and a tree. The token after each size is the blob hash:

pkg:npm/[email protected] — 1054 files, 1379.3 KB
└── package/
    ├── fp/
    │   ├── __.js  43B  QlKUJ6782LHNEISw-t4uX1hyH1TCZye4ShYOMIAghheg
    │   ├── _baseConvert.js  16.0K  QpGkoQltpQn5ZeTFxYQOnk8FWp-7yyeUQtyzdZXl48nA

package_file_contents

Read a single file from a package. Pass the hash printed next to an entry in package_files output. Returns up to 1 MB of UTF-8 text; binary files return metadata only.

package_file_grep

Search a single file from a package for lines matching a JavaScript regular expression, returning matches with line numbers (grep -n style). Each blob is fetched once and held in a process-wide cache, so repeated reads and greps of the same hash skip the network.

Authentication for organization-scoped tools

depscore works without credentials on the public server. The organizations, alerts, threat_feed, and package_files tools call Socket’s authenticated REST API, so they need a Socket API token.

How the server resolves a token depends on the transport:

• stdio mode reads one token at startup from the environment and uses it for every request. Set SOCKET_API_TOKEN. The server also accepts these aliases, in priority order: SOCKET_API_TOKEN → SOCKET_API_KEY → SOCKET_CLI_API_TOKEN → SOCKET_CLI_API_KEY → SOCKET_SECURITY_API_TOKEN → SOCKET_SECURITY_API_KEY. SOCKET_API_TOKEN is canonical; SOCKET_API_KEY is the alias most local setups already export. Because the process belongs to one user, this token is yours and scopes every tool to your account.
• HTTP mode scopes the organization tools to the caller, never to the server’s own token. Send your credential as an Authorization: Bearer header on each request. Which credential to send depends on whether the deployment runs OAuth. On an OAuth-enabled server, send an OAuth access token: every bearer token is validated through introspection, and a raw Socket API token is rejected with a 401 challenge no matter what it starts with. On a server without OAuth, send your raw Socket API token and it is used directly. Either way the server uses that per-request token for the Socket API calls it makes on your behalf. A shared deployment never answers organizations, alerts, threat_feed, or package_files with the operator’s data: when a request carries no token, those tools return the auth-required error. depscore alone may fall back to the server’s startup token, since package scores are the same for every caller.

When a token is missing, every affected tool returns the same message:

Authentication is required. Set SOCKET_API_TOKEN for stdio mode, or send your Socket API
token as an `Authorization: Bearer ` header (or connect through OAuth) in HTTP mode.

Generate a token from the Socket dashboard under API tokens, then export it before launching the server:

export SOCKET_API_TOKEN="your-socket-api-token"

Worked example: organization details and alerts

With SOCKET_API_TOKEN set, ask your assistant something like “show me the open critical alerts for my Socket org”. Under the hood the assistant chains two tools:

1. Discover the org slug. Call organizations (no arguments). The server reads your token, calls GET /v0/organizations, and returns the organizations your token can see. Pick the slug you want, e.g. acme-robotics.

2. Fetch alerts for that org. Call alerts with the slug and any filters:

   {
     "org_slug": "acme-robotics",
     "severity": "high,critical",
     "status": "open"
   }
   

   The server calls GET /v0/orgs/acme-robotics/alerts with the same token and returns the matching alerts plus pagination metadata. To page forward, pass the response’s endCursor back as cursor.

The same token scopes every org-scoped tool, so threat_feed and package_files work the moment organizations confirms which slug the token belongs to.

Adjusting tool usage via client rules

You can customize how the MCP server interacts with your AI assistant by editing your client’s rules file:

Example rule:

Always check dependency scores with the depscore tool when you add a new dependency. If the score is low, consider using an alternative library or writing the code yourself.

Claude Code Hook (Optional)

The repo ships an optional Claude Code hook that blocks high-risk packages before installation. When Claude Code runs an install command, the hook queries the public Socket MCP server at https://mcp.socket.dev/ and denies the install when the package’s supply chain score is below 20 (known malware, typosquats, high-risk supply chain signals). No CLI to install — copy the file and wire it up; the public server signs in via OAuth on first use.

Supported ecosystems and package managers:

Setup

Prerequisites: Node.js 24+.

1. Copy the whole dist/socket-gate directory into your hooks folder. The bundled socket-gate.cjs is self-contained, so it runs without any dependencies beside it.

   From a published install:

   mkdir -p ~/.claude/hooks
   cp -R node_modules/@socketsecurity/mcp/dist/socket-gate ~/.claude/hooks/
   

   From a checkout, build it first:

   pnpm run build
   mkdir -p ~/.claude/hooks
   cp -R dist/socket-gate ~/.claude/hooks/
   

2. Add to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "node ~/.claude/hooks/socket-gate/socket-gate.cjs"
          }
        ]
      }
    ]
  }
}

See hooks/socket-gate/README.md for the full reference.

How it works

The hook denies installation when supplyChain < 20, allows it otherwise — e.g. express/lodash/react (75–97) allow, browserlist (typosquat of browserslist, 15) and confirmed malware (0) block. Network, timeout, or parse errors all fail open, so a Socket outage will not block legitimate work.

Limitations

A best-effort guardrail, not a complete defense. Known gaps:

• Manifest edits + lockfile installs. If Claude edits a manifest directly (package.json, requirements.txt, Cargo.toml, Gemfile, go.mod, *.csproj) then runs a bare install (npm install, pip install -r requirements.txt, cargo build, bundle install, go mod tidy, dotnet restore), there is no package name on the command line to check.
• Package-manager invocations only. Direct downloads (curl | sh, wget), post-install scripts of already-accepted packages, and transitive dependencies are not re-checked.
• Indirect Claude paths. Sub-agents, MCP tools that shell out, and non-Bash tool calls are not covered unless the matcher is broadened.

Inspired by Jimmy Vo’s dependency hook.

Development

License

MIT