apify-mcp-server
- Type
- MCP server
- Scanned
- Nov 13, 2026
- Version
- fccd2d0965cc
- Files read
- 451
This server connects your AI assistant to Apify so it can search documentation, fetch web pages, run cloud web scrapers, and manage data or automations in your Apify account. The biggest risk is that hidden instructions on scraped websites could trick your AI into spending your Apify credits, running unwanted third-party tools, or sending private conversation details to outside sites, sometimes without asking for your confirmation.
What it can reach
- Secrets
- Reads your Apify API token from environment variables, request headers, or your saved Apify login file, and can forward optional payment credentials if provided.
- Files
- Reads its own bundled files and your local Apify login file (~/.apify/auth.json) when no API token is set, and does not read or edit other files on your computer.
- Network
- Connects to Apify servers and cloud scrapers, documentation search services (Apify, Crawlee, and Algolia), Segment and Sentry for usage and error reporting by default, and any websites fetched through Apify cloud tools.
- Other
- Starts cloud scraper runs and builds billed to your Apify account, reads stored datasets, and when optional management tools are enabled, can create, change, schedule, publish, or permanently delete resources in your Apify account.
Findings
-
LowProxied third-party MCP Server Actors can forge readOnlyHint to bypass user confirmation
When an Actor that runs its own MCP server is loaded as a tool source, getMCPServerTools copies tool.annotations directly from the remote Actor without checking whether the Actor is maintained by Apify or restricted in permissions. Although the README states that readOnlyHint is granted only to Apify-maintained Actors with limited permissions and all other Actor tools are marked destructive, any community MCP Actor can advertise readOnlyHint: true and openWorldHint: false. Clients such as Claude Desktop, Claude Code, and Cursor auto-approve read-only tools without asking the user, allowing prompt injection to silently invoke untrusted third-party Actor tools.
Fix: Override annotations on proxied Actor MCP tools unless the Actor is verified as an official Apify Actor with limited permissions, forcing readOnlyHint: false, destructiveHint: true, and openWorldHint: true for third-party Actors.
-
LowServer Actor tools compute readOnlyHint from batch input schema instead of HTTP API operations
src/tools/actors/actor_tools_factory.ts:255
When building dedicated tools for Server Actors (including default tools like apify--web-fetch and any selected Server Actor), getServerActorsAsTools sets readOnlyHint by calling isReadOnlyActor(actorInfo). That helper inspects only definition.input (the batch-run schema, which can be empty on a Server Actor) and ignores the HTTP OpenAPI schema in webServerToolInfo, such as mutating POST, PUT, PATCH, or DELETE methods or query/body parameters carrying custom headers or secrets. As a result, official Server Actors with state-changing HTTP endpoints or custom header support are still marked readOnlyHint: true and destructiveHint: false, causing MCP clients to skip confirmation prompts.
Fix: Require Server Actors to expose only safe HTTP methods (such as GET) and inspect their webServerSchema parameters before setting readOnlyHint: true.
-
LowSentry telemetry remains active when --telemetry-enabled false is passed as two CLI arguments
Before yargs parses CLI flags in stdio mode, instrument.ts checks only for the single combined argument '--telemetry-enabled=false', '--no-telemetry-enabled', or exact lowercase TELEMETRY_ENABLED==='false'. If a user opts out using space-separated '--telemetry-enabled false' in their MCP client args array (the same format used in the repository's own smoke tests and test helpers) or sets TELEMETRY_ENABLED=0 or FALSE, Segment telemetry turns off but Sentry stays enabled and continues sending session pings and crash reports to sentry.io.
Fix: Check process.argv for '--telemetry-enabled' followed by 'false' (case-insensitively) and parse process.env.TELEMETRY_ENABLED with parseBooleanOrNull so Sentry respects all documented opt-out forms.
-
LowWidget development server binds to 0.0.0.0 without authentication
In watch mode (pnpm run dev), esbuild's ctx.serve is called without specifying a host, which defaults to binding port 3226 on all network interfaces (0.0.0.0) and serving the src/web directory to anyone on the local network. Installed MCP server users never run this developer script.
Fix: Pass host: '127.0.0.1' to ctx.serve so the widget preview server listens only on loopback.
-
LowDeveloper evaluation harness auto-approves all Claude Code tools including Bash
evals/agent/claude_agent.ts:145
The developer evaluation runner enables the full Claude Code built-in toolset (including Bash and file edits) by default and supplies a canUseTool callback that unconditionally allows every tool call while passing API keys in the environment. When running merge-tier evaluations that fetch live third-party websites, prompt injection on a target page could execute shell commands on the runner machine. This code is never shipped or executed in a normal MCP installation.
Fix: Default mcpToolsOnly to true or restrict canUseTool to an explicit allowlist of non-shell tools during live-web evaluations.
-
LowDocker runtime image runs the server as root
The runtime stage of the Dockerfile (FROM node:24-alpine) never switches to the built-in non-root node user before launching dist/stdio.js. Users who run this MCP server via Docker run the Node.js process as root (UID 0) inside the container, granting unnecessary privileges if a runtime or dependency flaw is ever exploited.
Fix: Switch to the non-root user (USER node) in the runtime stage of the Dockerfile before ENTRYPOINT and ensure /app is readable by that user.
-
LowShell expansion in maintainer bug-triage skill when posting GitHub comments
.claude/skills/bug-triage/SKILL.md:130
This repository-only maintainer skill reads untrusted public GitHub issues and comments, drafts a Markdown reply (which commonly includes backticked code or $APIFY_TOKEN references), and instructs the agent to run gh issue comment with the reply inside double quotes in Bash after user approval. Because Bash expands backticks, $(...), and $VAR inside double quotes, approving a normal-looking Markdown draft can execute commands or leak shell environment variables into a public GitHub comment; end users who install the npm package or Docker image are not affected because .claude/skills is not included in the published package.
Fix: Pass the comment body over standard input using gh issue comment <number> --repo apify/apify-mcp-server --body-file - with a single-quoted heredoc so backticks and dollar signs in Markdown are not expanded by the shell.
