Nodgate: Approve Every MCP Tool Call Before It Runs
A popup on your Mac for every MCP tool call from Claude Desktop or claude.ai. Allow runs it, Block stops it, no answer blocks it.
$9
Works with the AI Agents you already use
Nodgate: Approve Every MCP Tool Call Before It Runs
Example session with this skill installed
Set up Nodgate and protect my "files" MCP server in Claude Desktop, so every tool call asks me first.
- Read your context and instructions
- Compiled the nodgate
Summary
Nodgate is installed and your files server is protected. From now on, every tool call Claude Desktop makes to it waits for your Allow or Block.
What I did
Set up (scripts/setup.sh): found Python 3.13 with Tk, created a private environment in the skill folder, installed Nodgate. Nothing global, no sudo.
2. Started the broker in its own Terminal window: nodgate broker
3. Listed your servers (names only, your API keys are never printed):
files not protected
web not a local server, cannot be protected
- Showed you the change first (dry run), and you said yes:
"files" in ~/Library/Application Support/Claude/claude_desktop_config.json (env and other keys stay as they are)
now: {"command": "npx", "args": ["-y", "@x/files"]}
after: {"command": ".../nodgate/app/.venv/bin/python", "args": ["-m", "nodgate.shim", "--name", "files", "--", "npx", "-y", "@x/files"]}
Nothing was written (dry run).
- Protected it:
Wrapped: files
Backup of the original config: .../claude_desktop_config.json.nodgate-backup
What happens now
| You do | Claude gets |
|---|---|
| Click Allow | The server's real result |
| Click Block or close the popup | Blocked by user (Nodgate), and the server never sees the call |
| Nothing for 45 s | Blocked by Nodgate: no decision was made in time |
| Broker not running | Blocked by Nodgate: the Nodgate broker is not running |
web is a remote connector, so Claude Desktop does not start it and it cannot be wrapped. For claude.ai connectors, see reference/claude-web.md.
Next steps
- Quit Claude Desktop with Cmd+Q and open it again so it loads the new config.
- Ask Claude to use any tool on
files: the popup appears above your other windows. - To undo:
nodgate uninstallputs the config back byte for byte.
Connects securely to your tools. The creator never sees your data.
About this skill
Claude can run MCP tools without asking you each time. Nodgate makes it ask.
When Claude Desktop, or a claude.ai custom connector, calls a tool on a server you protect, Nodgate holds the call and opens a window above all your other apps: which server, which tool, and the exact arguments. Click Allow and the call runs. Click Block and the server never sees it; Claude is told "Blocked by user (Nodgate)" and carries on.
It fails closed. No click within 45 seconds, the broker not running, a crash, a malformed message: all of these block. Nothing runs without an explicit Allow.
A working app, not a prompt
The skill sets up a small local Python application and walks your agent through it:
nodgate serverslists your Claude Desktop servers and which are protected. Names only: it never prints commands or API keys from your config.nodgate install --only NAME --dry-runshows exactly what would change in your config before anything is written. Your agent shows you that and waits for your yes.nodgate installprotects only the servers you name and keeps a byte-for-byte backup.nodgate uninstallputs the file back exactly as it was.- Several calls at once each get their own popup and their own answer.
- Every decision goes to a local audit log (time, server, tool, arguments, outcome), with fields named token, password, secret or api_key hidden.
- claude.ai support: a gateway in front of your MCP server, reached through a Cloudflare quick tunnel, behind a random secret path and an Origin check.
Tested properly
236 automated tests with real processes, real sockets and the official MCP SDK, including ten simultaneous calls, timeouts, cancellations, server crashes and 17 deliberately ambiguous messages. Checked against the real apps: Claude Desktop starts protected servers through Nodgate, allowed and blocked calls were run from a real Claude Desktop and a real claude.ai chat, and uninstall restored the config byte for byte (matching SHA-256).
What you get
- SKILL.md: set up, protect, verify and undo, with safety rules for the agent (dry run first, never wrap a server you did not name, never print secrets, no tunnel without asking)
- The Nodgate app (Python, one dependency: aiohttp), installed into a private environment inside the skill folder. Nothing global, no sudo
- Guides: how it works, protecting claude.ai, troubleshooting, and a demo server to try it safely first
Requirements
macOS, Python 3.11+ with Tk (the python.org installer includes it). For claude.ai only: cloudflared.
Not covered (yet)
Windows and Linux, Desktop Extensions (.mcpb), MCP resources and prompts (tool calls only), remote servers that need OAuth sign-in, "remember my answer" rules.
macOS only. No support for Windows/Linux. Does not intercept MCP resources or prompts. Requires manual Claude Desktop restarts. Requires Python 3.11+ with Tkinter.
How to use
- 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
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
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 today
Requires
Permissions
File Scopes
Shell: setup.sh creates a private Python environment inside the skill folder (no sudo, nothing global) and the agent runs the nodgate commands. Network: pip downloads aiohttp (and the official mcp SDK for the optional demo) from PyPI; the optional claude.ai mode opens a Cloudflare quick tunnel only when the user asks. Files: Claude Desktop's config is changed only by `nodgate install` after the user approves a dry run, with a byte-for-byte backup that `nodgate uninstall` restores; nodgate never prints env values or keys. ~/.nodgate holds the broker token, gateway secret and audit log (mode 0600). Env: NODGATE_UPSTREAM_TOKEN, optional, for upstream MCP servers that need a bearer token.