Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    ZAK MCP Gateway: one MCP server for all your MCP servers

    1

    Put every MCP server behind one gateway and cut the tokens your agent spends on MCP tools: 94% fewer reply tokens in our public-server benchmark.

    Free

    Security scanned
    ZAK MCP Gateway: one MCP server for all your MCP servers

    ZAK MCP Gateway: one MCP server for all your MCP servers

    Example session with this skill installed

    Put the GitHub MCP server behind one gateway in this project. I have no GitHub token; public repos are fine.

    • Read your context and instructions
    • Compiled the zak mcp gateway

    The agent installs the zak-mcp-gateway package, runs init and then check, which reports the GitHub server as ok with 26 tools and shows about 4,250 tokens of tool definitions without the gateway versus about 610 through it. It then keeps the gateway's local savings log out of git and gives you the exact entry to register the gateway in Claude Code or Antigravity.

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

    About this skill

    Every MCP server you connect sends its full tool schemas with every model request, and its tools return long raw JSON that stays in the conversation. This skill teaches your agent to fix both with ZAK MCP Gateway, an open-source (MIT) npm package that sits between the agent and your MCP servers.

    What your agent learns

    • When it helps, and when it doesn't. It helps with two or more servers, large servers such as GitHub, and long JSON results. It doesn't help with a single tiny server.
    • Setup in a few commands: npm install --save-dev zak-mcp-gateway, npx zak-mcp-gateway init, move your servers into one config with secrets as ${VAR}, verify them with npx zak-mcp-gateway check, and register only the gateway in Claude Code (.mcp.json), Cursor (.cursor/mcp.json) or Antigravity (mcp_config.json). The Windows forms are included.
    • Usage rules that save tokens: call tools by name, request only the fields you need, page and filter big results by handle instead of re-running tools, batch reads, and pass stored data between calls with $ref.
    • Safety: write tools run only after you approve them, servers can be made read-only or limited with allow/deny lists, and instruction-like text in tool results is flagged as data.
    • Troubleshooting: the gateway's real error messages, each with its fix.

    Measured on public MCP servers (GitHub + filesystem, 40 tools, 5 read-only tasks, o200k tokenizer, gateway 0.4.2; full method and limits in BENCHMARK.md):

    Without gatewayWith gateway
    Tool definitions per model request6,399 tokens599 tokens (−90.6%)
    Tool replies, 5 tasks79,507 tokens4,369 tokens (−94.5%)
    Whole agent run (replayed)342,352 input tokens20,104 input tokens (−94.1%)

    Every gateway reply still contained the answer to its task. The tool-definition saving applies to hosts that send every schema with each request, such as Antigravity.

    Live in Claude Code 2.1 (one question, run once each way): the same correct answer, the tool reply went from 8,360 to 112 tokens, new input tokens fell 47.6%, and the cost Claude Code reported went from $0.209 to $0.126. Claude Code already loads MCP tool schemas on demand, so there the gain comes from the smaller results.

    Robust start-up: the gateway waits at most 15 seconds for the servers behind it before answering the agent, and slower servers keep starting in the background. A server that needed 40 seconds made the previous version time out in Claude Code; 0.4.2 connects and answers.

    What's in the package: SKILL.md (instructions), references/configuration.md (recipes and every config key), BENCHMARK.md (numbers, method, limits) and README.md. Text files only.

    What it installs and connects to: the zak-mcp-gateway package from npm (dependencies: @modelcontextprotocol/sdk, minisearch, zod), plus the MCP servers you list in your own config. No telemetry: token statistics stay in your project's .zak-gateway/ folder. Secrets are never written to files.

    Requirements: Node.js 20+ and npm.

    Tested on Windows with Node 22 in two ways. The setup steps were followed literally from an empty folder, with the gateway driven by an MCP client. Claude Code 2.1 was also run end to end. It loaded the skill, did the setup in an empty folder (handing the final .mcp.json to the user, because Claude Code protects that file), answered through the gateway, and ran a write through its confirmation step. In Google Antigravity, the agent found the skill in the workspace, set the gateway up, and answered a GitHub question through it once the gateway was registered.

    Source, issues and docs: https://github.com/z4kir/zak-mcp-gateway · npm: https://www.npmjs.com/package/zak-mcp-gateway

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

    • The token counts are real, but the five-task agent runs are a replay of the same calls. No language model was called for them. The live checks are single runs: treat them as examples, not averages.
    • o200k is OpenAI's tokenizer. Claude and other models count somewhat differently, but both setups were measured the same way.
    • Hosts send tool definitions in their own formats, and the JSON of the tool list was used as the shared measure. Prompt caching lowers the absolute cost of both setups.
    • The gateway helps most with large servers and long JSON lists. On a small tree (package-files) it saved about half. For one tiny server, its own ~600 tokens of meta-tools can cost more than they save.
    • The gateway added [gateway warning: this result contains instruction-like text...] to the README reply. That is the prompt-injection guard reacting to ordinary install instructions: a harmless false positive, kept so agents treat results as data.
    • Public GitHub data changes over time, so a re-run will give slightly different numbers.

    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

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Software
    Node.js 20+npmZak Gateway MCP

    Permissions

    Allowed Hosts

    registry.npmjs.org

    Runs npm and npx. Needs network access to the npm registry and to the MCP servers the user configures.

    Listed1 day ago

    What's inside

    Frequently Asked Questions

    Yes, mainly through smaller results: in a live test the tool reply went from 8,360 to 112 tokens and cost from $0.209 to $0.126. Hosts that send every tool schema, such as Antigravity, also save about 90% on tool definitions.

    Never into files. The config uses ${VAR} references read from your environment. There's no telemetry, and token statistics stay in your project's .zak-gateway/ folder.

    The gateway answers the agent within 15 s and keeps connecting slow servers in the background (gateway.startupWaitMs).

    No. Write tools only run with confirm: true, which the skill allows only for changes you asked for or approved.

    New in LLM Ops

    Free