Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    MCP Connector Builder

    1

    Builds a Streamable HTTP MCP server, hardens the host, guards SSRF on user-supplied URLs, and verifies it with two independent clients.

    $21

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    MCP Connector Builder

    MCP Connector Builder

    Example session with this skill installed

    Build me an MCP server that exposes our internal pricing table read-only, and show me how to add it as a remote connector in Claude.

    • Read your context and instructions
    • Compiled the mcp connector builder

    Transport: Streamable HTTP on 127.0.0.1:8811 behind nginx. Skeleton with mcp<2 pinned, one read-only tool, per-IP 30/min plus 20/min per tool. Two-client check: probe_connector.py reports 1 tool, Inspector lists the same tool.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Builds an MCP server that other people's agents can actually use — and states plainly which part of "it works" was proven.

    What is inside

    • Transport decision + server skeleton. stdio vs Streamable HTTP, and a FastMCP skeleton with the SDK pin that matters (version 1.x — the 2.x line renamed FastMCP, so an unpinned install breaks every sample), stateless_http, json_response, a small request-body cap, loopback binding, and allowed_hosts that matches what nginx forwards.
    • Tool rules for a connector a third party can add. Read-only, no keys held by the connector, one config file for every changeable value, return data instead of instructions, honest tool names, call counts logged instead of payloads, and no "call any URL with any headers" passthrough.
    • Abuse controls before the first request. Per-IP limit in ASGI middleware (tools cannot see the caller's IP), per-tool sliding windows, a global hourly cap on outbound tools, and a kill switch that is a file checked by both the middleware and every tool. Plus the nginx limit_req_zone detail that trips everyone up: it must live in the http context, not inside a server block.
    • A fail-closed SSRF guard, standard library only, for any tool that fetches a URL a user handed it: scheme and port allowlist, every resolved address validated, the smart IP forms most filters miss (2130706433, 0x7f000001, 127.1, 0177.0.0.1), cloud metadata endpoints, pinning the validated resolution for that one request, and no redirects. Includes a 20-case offline self-test and the bypass list it must catch.
    • Hosting. A hardened systemd unit (dedicated user, ProtectHome, ProtectSystem=strict, empty capability set, memory/CPU caps), the nginx location block, and the nsenter command that proves the service cannot see your home directory, keys or session database.
    • Verification protocol. A stdlib wire-protocol probe (initialize → notifications/initialized → tools/list → tools/call) that needs no SDK and no Node.js, MCP Inspector as a genuinely independent second client, the security matrix to run (SSRF cases, redirect, DNS rebinding with a control case, oversized body, malformed input, rate-limit rain), and the four-layer check for publishing to an MCP registry.

    What it will not do. It will not host the service for you — opening a public route (nginx, DNS, firewall, TLS) is an operator action and the skill deliberately stops at the loopback service. It will not make a badly designed upstream API safe. It does not store credentials; a public connector built with it should hold none. A clean test run means the listed cases passed, not that the endpoint is secure.

    Tested before publishing. ssrf_guard.py --selftest passes 20/20 offline cases including the control case (which is what caught a real bypass in the first draft of the guard: the POSIX short form 127.1). The probe completed a real handshake against a live remote MCP server — initialize, then 27 tools listed — and the guard completed a real pinned HTTPS fetch of a public page while refusing both a loopback target and a cloud-metadata target.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.

    It does not host anything for you: opening the public route (nginx, DNS, firewall, TLS) stays an operator action. It does not make a badly designed upstream API safe, and it stores no credentials — a public connector built with it should hold none. Passing self-tests means the listed cases passed, not that an endpoint is secure.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 6 days ago

    Software
    Python 3.9+mcp package v1.x (to build)@modelcontextprotocol/inspector
    Optional
    Node.js 18+

    The two scripts use the standard library only. Building a server needs the mcp package v1.x; Node.js is only for the optional second client.

    Permissions

    File Scopes

    the
    project
    directory
    you
    build
    server
    in

    Network use is limited to what you point it at: one HTTPS request per probe to the MCP endpoint you name, plus one request to any URL you pass to the guard's fetch. No credentials are stored, read or forwarded.

    Listed6 days ago

    What's inside

    Frequently Asked Questions

    No. Both bundled scripts are standard-library only, so you can probe an existing server and test your SSRF guard without installing anything. The SDK (mcp v1.x) is needed only when you build a server.

    Run scripts/probe_connector.py against the endpoint with --expect-tool: a 200 with an empty tools/list is an empty server, not a healthy one. The failure table covers the usual causes next (421 Host header, handshake order, mcp 2.x installed).

    It makes it fail closed for a defined list of cases — 20 of them run offline, including the short/encoded IP forms and a control case. That is evidence for those cases, not a security guarantee, and the checklist tells you the live cases to add for your environment (redirects, DNS rebinding with a control).

    No. It reads first (inventory of running MCP services, ports, nginx blocks) and reuses the shape of an existing unit rather than forking tool lists. Everything it writes is new and named by you.

    New in Security