everything
modelcontextprotocol/servers/src/everything
- Type
- MCP server
- Scanned
- Nov 12, 2026
- Version
- e1dc4b18dbb9
- Files read
- 94
This is the official Model Context Protocol reference server designed to test tools, prompts, resources, sampling, elicitation, and transports for MCP client developers. If installed into an everyday coding agent, prompt injection can read credentials whose names miss get-env's keyword filter (such as DATABASE_URL or PGPASSWORD) and make outbound requests or reach internal services via gzip-file-as-resource.
What it can reach
- Secrets
- Reads all server process environment variables via get-env, masking values only when the variable name has an underscore-separated segment matching KEY, TOKEN, SECRET, PASSWORD, PASSWD, CREDENTIAL, AUTH, or COOKIE (or returning everything unmasked if GET_ENV_INCLUDE_SECRETS is enabled).
- Network
- Fetches arbitrary public http and https URLs via gzip-file-as-resource, and opens an unauthenticated HTTP server on port 3001 when started with the streamableHttp or sse transport.
- Files
- Reads only its own bundled documentation and package.json on disk, and can query the MCP client for configured workspace root URIs.
- Browser
- Can ask the MCP client to prompt the user to open a URL in their browser via URL-mode elicitation.
- Other
- Can request LLM completions (sampling) and interactive user forms (elicitation) from MCP clients that declare those capabilities.
Findings
-
MediumIncomplete credential masking in get-env leaves common secrets exposed
src/everything/tools/get-env.ts:71
The get-env tool is marked read-only (which many AI clients run without asking for confirmation) and splits variable names only on underscores to redact 8 specific words (KEY, TOKEN, SECRET, PASSWORD, PASSWD, CREDENTIAL, AUTH, COOKIE). Secrets stored in standard variables without those standalone segments—such as DATABASE_URL, REDIS_URL, PGPASSWORD, MYSQL_PWD, GH_PAT, SSH_PASSPHRASE, or camelCase names like npm_config__authToken—are returned in plain text to the AI model, where prompt injection could read them and exfiltrate them via gzip-file-as-resource.
Fix: Use an explicit allowlist of safe environment variables, or additionally redact URL values containing credentials, camelCase boundaries, and patterns like *PASSWORD*, *PWD, *PAT, *PASSPHRASE, *_URL, *_URI, and *_DSN, and set readOnlyHint to false.
-
LowSSRF filter in gzip-file-as-resource can be bypassed via DNS rebinding or NAT64
src/everything/tools/gzip-file-as-resource.ts:329
Before fetching a URL, gzip-file-as-resource resolves the hostname to block loopback, private LAN, and cloud metadata IP addresses, but fetch() resolves the hostname a second time when connecting. If an AI agent is tricked into passing an attacker-controlled domain whose DNS answer switches from a public IP to 127.0.0.1 or 169.254.169.254 between the two lookups (DNS rebinding), or an IPv6 NAT64 address under 64:ff9b::/96 (which is omitted from the blocklist), the server fetches the internal or cloud metadata response and returns its gzipped content to the agent.
Fix: Pin the validated IP address in a custom HTTP dispatcher so the connection uses the exact address that passed validation, block 64:ff9b::/96 in NON_PUBLIC_ADDRESSES, or restrict GZIP_ALLOWED_DOMAINS to trusted hosts.
-
LowHTTP transports have no authentication and grant CORS access to any localhost port
src/everything/transports/streamableHttp.ts:148
When started in streamableHttp or sse mode, the server exposes all MCP tools over HTTP without any password or token check. Any other user or process on the same computer can connect to 127.0.0.1:3001, any web page served from another localhost port (such as a local preview of untrusted HTML) is granted cross-origin access via CORS, and running the Docker image in HTTP mode with -p 3001:3001 (without 127.0.0.1:) exposes the unauthenticated server to the network because the Dockerfile sets HOST=0.0.0.0.
Fix: Require a secret Bearer token on HTTP endpoints (/mcp, /sse, and /message), restrict allowed CORS origins to an explicit configured origin rather than all localhost ports, and always publish container ports on 127.0.0.1 only.
-
Lowtrigger-url-elicitation allows non-HTTP URI schemes such as javascript: and file:
src/everything/tools/trigger-url-elicitation.ts:31
The trigger-url-elicitation tool validates its url parameter only with z.string().url(), which accepts any syntactically valid URI scheme including file:, javascript:, data:, and custom desktop app handlers like vscode:. If prompt injection tricks the agent into calling this tool, an MCP client that hands elicited URLs to the operating system's URL opener without strict scheme filtering could launch local files or desktop protocol handlers if the user accepts.
Fix: Restrict the url schema in TriggerUrlElicitationSchema to https: (and optionally http: for localhost testing) before sending the elicitation request.
-
LowUnbounded steps in trigger-long-running-operation can flood notifications and exhaust memory
src/everything/tools/trigger-long-running-operation.ts:9
Unlike other tools in this server that cap counts and sizes, trigger-long-running-operation places no upper limit on steps or duration. If tricked into calling it with duration: 0 and a huge steps count (such as 100000000), the server runs a tight loop sending millions of progress notifications, pegging the CPU and, on 2025-era Streamable HTTP sessions, filling the unbounded InMemoryEventStore until the server crashes out of memory.
Fix: Validate duration and steps as positive numbers with reasonable upper bounds (for example, steps as a positive integer up to 100 and duration up to 300 seconds).
